[CTDB and GFS] test: scenario and some results

[Ignacio Coupeau]

I tested the SAMBA cluster CTDB with GFS6.1. The main purpose of this
test was the LB ability of the CTDB.

The numeric results are very poor with concurrence (as usual with GFS).
The load balancing issue: well for CTDB, bad for my hardware.

I. Testing scenario
------------------------
- LB: Foundry ServerIron XL
	Virtual IP (VIP): X.Y.Z.185

- 3 nodes (2xXeon IBM 346)
	IP: X.Y.Z.181-183
	RH AS4U4
	GFS6.1 with lock_dlm
	samba_3_0_ctdb (070709)
	ctdb (070709)
	private network (eth1): Speed: 1000Mb/s FD
	public network (eth0):  Speed: 100Mb/s FD
	netbios name all nodes: smb-ctdb

-SAN:
	fabric: SanBox16
	data:	SUN3510 for GFS
		IBM DS4300 for EXT3

- Samba configuration:
	----
	...
	security = domain
	password server = pdcXYZ
	...
	locking posix = no
	...
	bind interfaces only = yes
	interfaces = eth0
	...
	clustering = yes
	#idmap backend = tdb2
	private dir = /usr/local/etc/users_1/clu
	use mmap = no
	posix locking = no
	-----

	Note1: Volker suggested me "posix locking = no";
	improve ~50% the results :-)

	Note2: for join the server to a domain, a patch
	for source/passdb/secrets.c

-------------------------------------------
II. Functional results
-------------------------------------------
1. The LB was a bit a mess, basically because the ServerIron I used
can't test the samba at L7 level (not in the ServerIron's well know
ports). So, if the eth0 interface is up, the traffic is balanced also to
the stopped samba, and the things goes very badly.

However, if I perform an "ifdown eth0" in one of the samba nodes, the LB
goes very well.

2. The service \\smb-ctdb\CLU01  are accessed from XP WS (map network
drive) and smbclient without any problem.

With XP the operation from an application performs well in a fail-over
(like a ms-word editing for example); and, when the interface comes
back up, only the window may clog a bit because the ServerIron is a bit
lazy.

-------------------------------------------
III. Scripts for smbtorture (SAMBA_4_0 070711)
-------------------------------------------

EXT3:
./smbtorture //smb-324/LOCAL -Utest%test RAW-BENCH-LOCK  --num-progs=32 -t60

GFS:
./smbtorture //smb-ctdb/CLU01 -Utest%test RAW-BENCH-LOCK
--unclist=unclist-3n.txt --num-progs=32 -t60

./smbtorture //smb-ctdb/CLU01 -Utest%test RAW-BENCH-LOCK
--unclist=unclist-2n.txt --num-progs=32 -t60

unclist-3n.txt
----------------
//X.Y.Z.181
//X.Y.Z.182
//X.Y.Z.183

unclist-2n.txt
----------------
//X.Y.Z.181
//X.Y.Z.182


-------------------------------------------
IV. The results
-------------------------------------------

EXT3 samba-3.24
---------------
1-node ops average = 8619 ops

GFS
--------------------------------
1-node ops average = 7401 ops
2-node ops average =  204 ops
3-node ops average =  469 ops

Note1: the smbtorture runs in a linux other than cluster nodes.
Note2: there are a high variability in the results, but in the same
        ratio.

-- 
________________________________________________________
Dr. Ignacio Coupeau
Systems and Network Services Director
IT Services
University of Navarra           http://www.unav.edu/
Pamplona, SPAIN                 http://www.unav.es/SI/