Re: [PATCH] vmscan: raise the bar to PAGEOUT_IO_SYNC stalls

[Mel Gorman]

On Fri, Jul 30, 2010 at 03:17:35PM +0200, Andrea Arcangeli wrote:
> On Wed, Jul 28, 2010 at 03:17:05PM +0800, Wu Fengguang wrote:
> > Fix "system goes unresponsive under memory pressure and lots of
> > dirty/writeback pages" bug.
> > 
> > 	http://lkml.org/lkml/2010/4/4/86
> > 
> > In the above thread, Andreas Mohr described that
> > 
> > 	Invoking any command locked up for minutes (note that I'm
> > 	talking about attempted additional I/O to the _other_,
> > 	_unaffected_ main system HDD - such as loading some shell
> > 	binaries -, NOT the external SSD18M!!).
> > 
> > This happens when the two conditions are both meet:
> > - under memory pressure
> > - writing heavily to a slow device
> > 
> > OOM also happens in Andreas' system. The OOM trace shows that 3
> > processes are stuck in wait_on_page_writeback() in the direct reclaim
> > path. One in do_fork() and the other two in unix_stream_sendmsg(). They
> > are blocked on this condition:
> > 
> > 	(sc->order && priority < DEF_PRIORITY - 2)
> > 
> > which was introduced in commit 78dc583d (vmscan: low order lumpy reclaim
> > also should use PAGEOUT_IO_SYNC) one year ago. That condition may be too
> > permissive. In Andreas' case, 512MB/1024 = 512KB. If the direct reclaim
> > for the order-1 fork() allocation runs into a range of 512KB
> > hard-to-reclaim LRU pages, it will be stalled.
> > 
> > It's a severe problem in three ways.
> 
> Lumpy reclaim just made the system totally unusable with frequent
> order 9 allocations.

Yes, it's very disruptive and has been for a while. It was not much of a
problem when resizing the static hugepage pool but is a disaster for
transparent huge pages.

> I nuked it long ago and replaced it with mem
> compaction. You may try aa.git to test how thing goes without lumpy
> reclaim. I recently also started to use mem compaction for order 1/2/3
> allocations as there's no point not to use it for them, and to call
> mem compaction from kswapd to satisfy order 2 GFP_ATOMIC in
> replacement of blind responsiveness-destroyer lumpy.
> 

A full-scale replacement is overkill but I can see why it would be done
in the short-term. There are times when lumpy reclaim is still needed -
specifically when the allocation failure is due to a lack of memory rather
than fragmentation. There will also be cases where compaction can't work
because there are too many movable pages  to move into too few pageblocks.

> Not sure why people insists on lumpy when we've memory compaction that
> won't alter the working set and it's more effective.
> 

Compaction is preferred, no doubt about it but lumpy reclaim cannot be
dismissed. I know lumpy reclaim is too disruptive and Kosaki noticed the same
and it's currently doing some pretty stupid things. There are a few ideas
knocking around publicly on how to reduce its impact while increasing its
effectiveness. I have a few old ideas knocking around as well that I just
need the time to get around to. I hope to get at it after the fuss over
writeback is addressed.

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/