Re: LFS (was: New FSs in 3.4?)

[Hannah Schroeter <hannah@schlund.de>]

Hello!

On Fri, Aug 29, 2003 at 08:50:21AM -0700, Claus Assmann wrote:
>[...]

>> that would really surprise me.  lots of fsync (hello mta) should drive lfs 
>> into the floor i would have expected.

>The trick is that you write just one logfile instead of writing
>little files all over the disk. An fsync() is just an update to the
>logfile, hence you are limited by the bandwith, not by the transaction
>rate.

>Just a data point: using JFS on Linux sendmail 9 can relay more
>than 100 msg/s, while a conventional FFS can only achieve 30 msg/s.

I thought that the great difference between JFSs (in general) and
LFSs (in general) is that LFSs are *only* the log, while JFSs are
a log plus more traditional structures. So a *full* fsync() would
be an append to the log on a LFS, while it'd be an append to the log,
then an update in the other structures on a JFS. Of course, you
could implement fsync() faithfully by doing only the append to the log
synchronously and deferring the update of traditional structures,
but then, operating at *that* throughput rate with an JFS would be
like living on debt (more and more deferred work still to do).

>[...]

Kind regards,

Hannah.
-- 
  Hannah Schröter            Entwicklung       hannah@schlund.de
  Bei Schlund + Partner AG   Brauerstraße 48   D-76135 Karlsruhe
This specification allows any of these approaches.  Solving the
Halting Problem is considered extra credit. (RFC 3028)