[lttng-dev] [-stable 3.8.1 performance regression] madvise POSIX_FADV_DONTNEED

[Mathieu Desnoyers]

* Dave Chinner (david at fromorbit.com) wrote:
> On Wed, Jul 03, 2013 at 10:53:08AM -0400, Jeff Moyer wrote:
> > Mel Gorman <mgorman at suse.de> writes:
> > 
> > >> > I just tried replacing my sync_file_range()+fadvise() calls and instead
> > >> > pass the O_DIRECT flag to open(). Unfortunately, I must be doing
> > >> > something very wrong, because I get only 1/3rd of the throughput, and
> > >> > the page cache fills up. Any idea why ?
> > >> 
> > >> Since O_DIRECT does not seem to provide acceptable throughput, it may be
> > >> interesting to investigate other ways to lessen the latency impact of
> > >> the fadvise DONTNEED hint.
> > >> 
> > >
> > > There are cases where O_DIRECT falls back to buffered IO which is why you
> > > might have found that page cache was still filling up. There are a few
> > > reasons why this can happen but I would guess the common cause is that
> > > the range of pages being written was in the page cache already and could
> > > not be invalidated for some reason. I'm guessing this is the common case
> > > for page cache filling even with O_DIRECT but would not bet money on it
> > > as it's not a problem I investigated before.
> > 
> > Even when O_DIRECT falls back to buffered I/O for writes, it will
> > invalidate the page cache range described by the buffered I/O once it
> > completes.  For reads, the range is written out synchronously before the
> > direct I/O is issued.  Either way, you shouldn't see the page cache
> > filling up.
> 
> <sigh>
> 
> I keep forgetting that filesystems other than XFS have sub-optimal
> direct IO implementations. I wish that "silent fallback to buffered
> IO" idea had never seen the light of day, and that filesystems
> implemented direct IO properly.
> 
> > Switching to O_DIRECT often incurs a performance hit, especially if the
> > application does not submit more than one I/O at a time.  Remember,
> > you're not getting readahead, and you're not getting the benefit of the
> > writeback code submitting batches of I/O.
> 
> With the way IO is being done, there won't be any readahead (write
> only workload) and they are directly controlling writeback one chunk
> at a time, so there's not writeback caching to do batching, either.
> There's no obvious reason that direct IO should be any slower
> assuming that the application is actually doing 1MB sized and
> aligned IOs like was mentioned, because both methods are directly
> dispatching and then waiting for IO completion.

As a clarification, I use 256kB "chunks" (sub-buffers) in my tests, not
1MB. Also, please note that since I'm using splice(), each individual
splice call is internally limited to 16 pages worth of data transfer
(64kB).

> What filesystem is in use here?

My test was performed on ext3 filesystem, that was itself sitting on
raid-1 software raid.

Thanks,

Mathieu

> 
> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david at fromorbit.com

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com