GFS ping_pong test results

Abhijith Das adas at redhat.com
Tue Aug 14 17:08:05 GMT 2007


Hi,
I saw a few posts (a few weeks old) with ping_pong test numbers on GFS1 and RHEL4.
The cluster-wide posix lock performance in RHEL4+GFS1 is not as good as it is 
in RHEL5+GFS1. 

How to setup the cluster:
http://sources.redhat.com/cgi-bin/cvsweb.cgi/~checkout~/cluster/doc/usage.txt?rev=1.39&content-type=text/plain&cvsroot=cluster

Below are the numbers we came up with.
gfs_controld limits plocks by default in order to not flood the network. You can 
allow unlimited plocks by starting gfs_controld with the -l0 option. 
'gfs_controld -l0'

Are these numbers comparable to gpfs?

Thanks,
--Abhi
                              1 node     2 nodes       3 nodes    4 nodes

4 node cluster
ping_pong test.dat         2550-2570   2070-2100     1780-1800    900-910
ping_pong test.dat -rw     2510-2560       40-45         25-30         19
ping_pong test.dat -rwm    2550-2570       70-90         30-35         23

3 node cluster
ping_pong test.dat         3100-3200   2400-2500     1350-2050 [*]
ping_pong test.dat -rw     3100-3200       30-45 [+]     20-30
ping_pong test.dat -rwm    3100-3200       50-90         28-32

2 node cluster
ping_pong test.dat         3700-3800   2700-2800
ping_pong test.dat -rw     3700-3800       35-45 [+]
ping_pong test.dat -rwm    3700-3800       55-70

1 node cluster
ping_pong test.dat         8300-8500
ping_pong test.dat -rw     6300-6500
ping_pong test.dat -rwm    7300-7700

- Tests ran with N of 4 and 5.

- The 3 node cluster had much less consistent results in general.

[*] Sometimes this number is fixed around 1350, other times fixed around 2050,
    don't know why the difference.
[+] This number I was seeing around 200 in the past.



More information about the samba-technical mailing list