On Mon, Mar 28, 2011 at 10:44:45AM +0800, Wu Fengguang wrote: > On Sat, Mar 26, 2011 at 07:05:44AM +0800, Jan Kara wrote: > > And actually the NFS traces you pointed to originally seem to be different > > problem, in fact not directly related to what balance_dirty_pages() does... > > And with local filesystem the results seem to be reasonable (although there > > are some longer sleeps in your JBOD measurements I don't understand yet). > > Yeah the NFS case can be improved on the FS side (for now you may just > reuse my NFS patches and focus on other generic improvements). > > The JBOD issue is also beyond my understanding. > > Note that XFS will also see one big IO completion per 0.5-1 seconds, > when we are to increase the write chunk size from the current 4MB to > near the bdi's write bandwidth. As illustrated by this graph: > > http://www.kernel.org/pub/linux/kernel/people/wfg/writeback/dirty-throttling-v6/4G/xfs-1dd-1M-8p-3927M-20%25-2.6.38-rc6-dt6+-2011-02-27-22-58/global_dirtied_written-500.png Which is _bad_. Increasing the writeback chunk size simply causes dirty queue starvation issues when there are lots of dirty files and lots more memory than there is writeback bandwidth. Think of a machine with 1TB of RAM (that's a 200GB dirty limit) and 1GB/s of disk throughput. Thats 3 minutes worth of writeback and increasing the chunk size to ~1s worth of throughput means that the 200th dirty file won't get serviced for 3 minutes.... We used to have behaviour similar to this this (prior to 2.6.16, IIRC), and it caused all sorts of problems where people were losing 10-15 minute old data when the system crashed because writeback didn't process the dirty inode list fast enough in the presence of lots of large files.... A small writeback chunk size has no adverse impact on XFS as long as the elevator does it's job of merging IOs (which in 99.9% of cases it does) so I'm wondering what the reason for making this change is. Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
