On Fri, Feb 19, 2010 at 11:07:30AM +1100, Anton Blanchard wrote:

Hi,

quoted

The patch below sets a smaller value for RECLAIM_DISTANCE and thus enables
zone reclaim.

I've no problem with the patch anyway.

FYI even with this enabled I could trip it up pretty easily with a multi
threaded application. I tried running stream across all threads in node 0. The
machine looks like:

node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
node 0 free: 30254 MB
node 1 cpus: 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
node 1 free: 31832 MB

Now create some clean pagecache on node 0:

# taskset -c 0 dd if=/dev/zero of=/tmp/bigfile bs=1G count=16
# sync

node 0 free: 12880 MB
node 1 free: 31830 MB

I built stream to use about 25GB of memory. I then ran stream across all
threads in node 0:

# OMP_NUM_THREADS=16 taskset -c 0-15 ./stream

We exhaust all memory on node 0, and start using memory on node 1:

node 0 free: 0 MB
node 1 free: 20795 MB

ie about 10GB of node 1. Now if we run the same test with one thread:

# OMP_NUM_THREADS=1 taskset -c 0 ./stream

things are much better:

node 0 free: 11 MB
node 1 free: 31552 MB

Interestingly enough it takes two goes to get completely onto node 0, even
with one thread. The second run looks like:

node 0 free: 14 MB
node 1 free: 31811 MB

I had a quick look at the page allocation logic and I think I understand why
we would have issues with multple threads all trying to allocate at once.

- The ZONE_RECLAIM_LOCKED flag allows only one thread into zone reclaim at
  a time, and whatever thread is in zone reclaim probably only frees a small
  amount of memory. Certainly not enough to satisfy all 16 threads.

- We seem to end up racing between zone_watermark_ok, zone_reclaim and
  buffered_rmqueue. Since everyone is in here the memory one thread reclaims
  may be stolen by another thread.

You're pretty much on the button here. Only one thread at a time enters
zone_reclaim. The others back off and try the next zone in the zonelist
instead. I'm not sure what the original intention was but most likely it
was to prevent too many parallel reclaimers in the same zone potentially
dumping out way more data than necessary.

I'm not sure if there is an easy way to fix this without penalising other
workloads though.

You could experiment with waiting on the bit if the GFP flags allowi it? The
expectation would be that the reclaim operation does not take long. Wait
on the bit, if you are making the forward progress, recheck the
watermarks before continueing.

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab