[Linux-cluster] GNBD client, memory starvation

[Jeff Hardy]

I am having a few issues with memory exhaustion on gnbd clients when
writing large files to a gnbd server that re-exports ATAoE storage.
These are files 6 or 7 GB or larger.  If I export a SCSI disk from that
same gnbd server I do not have issues, so it leads me to believe the
slower ATAoE storage relative to local disk is causing the cache to fill
up on the gnbd client.

On the gnbd server, I export some ATAoE storage that has been chunked up
by LVM:
gnbd_export -c -d /dev/san00/lv00 -e lv00

And on the client:
gnbd_import -i 10.0.0.1 -n

Can any one suggest some proc settings to fiddle with, or other things
to try?  The gnbd server and clients are both FC4.  Just one client, so
no fencing or anything else.  I have tried fiddling with different
values in a couple of proc settings, and it seems to have helped a bit.
In fact, I could not reliably write large files to the exported SCSI
disk either, until I made these changes:

echo 5 > /proc/sys/vm/dirty_background_ratio
echo 5 > /proc/sys/vm/dirty_ratio

My testwrite script is a Perl script that just writes 1's to a file.