[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: writeback on /tmp and laptop?



Hi,

On Fri, 2003-03-14 at 15:45, Matthias Andree wrote:

> >> Wwould the Linux 2.5 (or maybe even 2.4, but I don't dare to dream of
> >> that) version be able to cope with -W1 through using ordered tags or
> >> flush cache commands?
> >
> > In theory, yes, but the code is not integrated.
> 
> Well, whom do I tell that I'd really like to see that in Linux 2.6?

lkml in general; Jens Axboe and the journaling fs folks in particular.

> > To some extent, yes.  The trouble is, once you add in LVM or software
> > raid, all the hardware help disappears since you now have the added
> > problem of having to synchronise over multiple disks.  A cache flush is
> > a much easier primitive to push down through the logical-device stack,
> > though.
> 
> Well, yes, but the RAID or LVM or EVMS layer should be aware of what our
> logical block unfurls to -- so are these subsystems lacking the
> necessary interfaces?

It's not that --- it's that the interface suddenly becomes useless.

Barrier ops have the nice property that you can push them right down to
the hardware without breaking the IO pipeline, and the disks will
observe your write ordering constraints.

Once you've got more than one disk, you simply cannot do that.  There
just isn't any mechanism to tell one disk "here's a write, but don't
complete it before that other disk over there has done all previous
writes."  So you end up stalling the IO pipeline anyway, and keeping the
IOs in the OS queues until you get an ACK that the barrier IO has been
completed.

So sure, you _can_ implement it purely in software if you've got logical
devices in the mix; you just don't really benefit from it at that point.

Cheers,
 Stephen





[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]