Andi Kleen wrote:
>
> On Wed, Jun 07, 2000 at 06:49:14PM -0700, Rajagopal Ananthanarayanan wrote:
> > Results of write performance tests:
> > ----------------------------------
> >
> > Sequential write using lmdd, file size is ~209MB on a 2 CPU system with
> > total memory of 64M. The experiment is run over varying write-size from
> > 1K to 1024K, 3 times for each block-size. This shows
>
> >
> > 1. Ext2 is well tuned (hardly any variation in 3 runs of each blocksize)!
>
> Please not that 2.3 itself has significant performance regressions for
> huge bulk writes (there were several threads on linux-kernel about that).
> Partly the still broken page cache balance is probably to blame, for
> other things the elevator (Jens Axboe's per device elevator patches seem
> to cause a huge speedup)
Yep, I lurk in linux-mm to garner some of this news; but thanks for the heads
up.
> With tuning 2.3.99pre2, an very old kernel, you might be duplicating
> work that others already did.
No, the recent changes were to tune XFS itself rather than Linux VM.
Hopefully all the recent tuning in 2.4.0+ will be beneficial for XFS
as it is for ext2 ...
We have one more set of changes to go in the write path so that
pagebuf/kiobufs are used to really cluster the writes onto disk ...
Right now, this is done through kflushd/ll_rw_block/elevator. Some of
these clusters can be large (thousands of pages long) to a single extent
(contiguous blocks on disk). So (a) the clusters don't need to be "discovered"
by an elevator-like algorithm (b) kiobuf based I/O will avoid processing
thousands of buffer-heads.
--
--------------------------------------------------------------------------
Rajagopal Ananthanarayanan ("ananth")
Member Technical Staff, SGI.
--------------------------------------------------------------------------
|