Re: LRU frequency

On Wed, Jul 20, 2005, Lucas Brasilino wrote:

> Since you mentioned, I looked event.c code and I saw that
> every "event functions" called within eventRun() is unregistered.
> So functions like storeMaintainSwapSpace() must be re-registered.
> Why not add a flag to struct ev_entry that tells if an event should
> be unregistered or not ? It will save some eventAdd() calls...

They're relatively cheap, especically at the frequency they're
being called at for LRU maintainence.

> >Actually, it means that Squid cannot delete old objects fast
> >enough to make room for new ones.
>
> Hmm should be due blocking unlink() behavior.
> What about rising unlinkd priority with nice(2) in extreme
> cases of swap out needing? The bad side is that unlinkd should
> be root setuid.

Well, there's certainly an upper limit of how fast your
filesystem can handle unlink()ing lots of files.
I'm pretty sure modern ones are doing it asynchronously
(ie without some implicit FSYNC afterwards to make sure
the directory metadata is happy).

> >What is the cache size and request rate of these
> >caches that quit due a full cache directory?
>
> Sorry, I don't remember. It has some time,
> so nevermind.

It would certainly be nice to know.

Adrian
Received on Wed Jul 20 2005 - 18:52:13 MDT