On Tue, 2002-03-19 at 11:36, Andi Kleen wrote:
> Hallo Steve,
> On Tue, Mar 19, 2002 at 11:23:05AM -0600, Steve Lord wrote:
> > Some reorganization of the memory allocation for the xfs log
> > buffers. This has two affects, it stops us from using a 64K
> > chunk of memory for the 32K buffer - we now use 32K chunks.
> > The other is to make the log writes start on page boundaries
> > which will help raid5 - and everyone else probably.
> Does that imply that a higher logbufs= value is recommended now, or
> does XFS need more changes to keep more log in memory ?
No, this was just a couple of inefficiencies, the in core log
buffers consisted of a control structure (512 bytes) and 32K of data,
all allocated in one chunk. End result was we chewed up 64K for
each, and all log writes started 512 bytes into a page. The
latter was one of the causes of raid5 cache thrashing, but there
is more to do on that problem.
p.s. Andi you should look into the next checkin - it might fix stuff
Steve Lord voice: +1-651-683-3511
Principal Engineer, Filesystem Software email: lord@xxxxxxx