[Linux-cluster] Tuning GFS and fnctl locks

["Peter Shearer" <pshearer lumbermens net>]

Title: Message

Hi, All --

 

I'm playing around with FC4 (x86_64smp - Dell PowerEdge 1850 hardware) and the GFS rpms included in the distribution.  Couple of questions:

 

I know that there's a list of tunable options given by gfs_tool gettune <mountpoint>.  Is there a list somewhere of what these options do?

 

Is there a way to optimize GFS with regards to opening/closing files with fnctl locks?  My app does a lot of writing with fnctl locks (file-based database application), and it really drags on GFS when running distributed locks.  I've posted about this in the past; I'm just now getting around to looking at GFS again.  I haven't seen a problem with regards to actual I/O throughput over my iSCSI SAN - my app under GFS typically will do 30 kB/s, and I have been able to push 100+ MB/s through the SAN - it's just slowing down significantly when the files are opened with a fnctl lock.  The top command gives the following:

 

  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
 5996 root      18   0  4268 2700 1564 R 39.8  0.1   0:01.20 runcobol
 3201 root      10  -5     0    0    0 S 22.9  0.0   8:20.86 dlm_astd
 4285 root      10  -5     0    0    0 S 15.9  0.0   0:43.59 lock_dlm2
 4284 root      10  -5     0    0    0 S 13.9  0.0   0:44.91 lock_dlm1

I've seen dlm_astd and the two lock processes go higher than this in CPU time, sometimes up to 30%+.

 

For some reason I remember reading somewhere that I should disable hyperthreading when using GFS.  I can't seem to find where I read that statement.  Am I'm just going insane, or is that really the recommendation?

 

We'd like to see three or more of the machines in the mix; currently there is only one.  Any suggestions would be greatly appreciated.

 

Thanks!

 

--Peter