[Linux-cluster] GFS over GNBD freezing problems

[andremachado]

Hello,
I am trying to spot the cause of a seemingly random freezing of GFS over
GNBD.
Configured 3 gnbd servers. Only one actively mounted.
A fourth machine is the mounting point.
When writing files, many times it hard freezes the filesystem. 
Even using gnbd_import -RO cannot release the resource.
Only resetting and gfs_fsck after, with many fs corruption.
Tried to gnbd_export -t 600 from the gnbd server.
Tried an exhagerated clvmd -t 1800 roundtrip.
Tried to export raw partition /dev/hda3 (faster).
Tried to export logical volume (slower).

I discovered some interesting things.
If gfs is mounted at the local machine, no problems, even with big
files.
The problems arise when gfs over gnbd.
If the writing is SLOW, no problems.
Used an internet download throttled by wget.
But if fast writing from disk, or from CD or from another lan machine,
the filesystem hangs.
Also, if writing from disk, but from directories not previously READ,
then it is much more *likely* to succeed.
But if previously copied / read to another directory, and THEN writing
to gfs mounting point, the problem is very likely to appear at the
beginning.
I saw that a 150 MB file previously copied to another directory was
"presumably" written to gfs in a snap, then when writing the next small
files, the filesystem hanged.
gnbd_export without cache, as per documentation.
It seems that cached files are too fast for gfs.
The whole cluster is using regular tcp ethernet at the same segment.
There is no private net for the cluster.
It seems that the gfs is missing some configuration to be more slow
connection or lag tolerant.
But the timeout configurations I tried had no effect.
It seems there is no sync mount option for gfs_mount.
What configuration am I forgetting?

Regards.
Andre Felipe Machado



[0]
http://www.techforce.com.br/index.php/news/linux_blog/red_hat_cluster_suite_debian_etch