[Linux-cluster] GFS performance problem

Riaan van Niekerk riaan at obsidian.co.za
Wed Oct 25 08:45:21 UTC 2006



FM wrote:
> Hello,
> 
> Here is my setup :
> RedHat Enterprise 4 (update 3).
> 5 web servers connected to a 600 GB  GFS (noatime) on a SAN.
> On the GFS : all web sites ROOTS and httpd logs files
> All web servers are writing to the same log. No problem here.
> 
> The prob is with rsync. When writing on the GSF it is very very slow !
> (40 % slower when we are lucky :) )
> 
> lots of question :
> *) could it be because if I have 1 GFS for the heavy write on the log
> and another GFS for the websites' files ?
> *) expect for the noatime, is there other technic to speed up GFS ?

a) we have increased the number of GFS locks cached to bring them closer 
in line with the number of locks in use:
echo "200000" > /proc/cluster/lock_dlm/drop_count
(not sure how much of a performance increase this got us)

b) We have disabled quotas
gfs_tool settune /mnt/san quota_account 0
We saw a 3 - 5 increase in performance

c) if you are using lots of small files, look into section 5.8 in the 
GFS manual, Data Journaling. However,
- I dont know how to change this for existing data on a GFS
- I have asked on the mailing list, and no-one seems to be using it.

> *) Can several webserver access EXT3 FS (read only) when only on other
> server have RW access to it ?

no. the RO server will get confused when things change from under it, 
since it is not expecting things to change

> *) is there a options to tune rsync when using GFS ?
> *) we are using DLM as the locking system. All servers are connected
> with Gb RJ45. Is DLM using the network to manage the lock. And if it is
> the case, could my problem come from the network latency ?

DLM is using the network, yes. not sure about latency. We use GB 
ethernet with RJ45/CAT5+ and have not had any problems related to DLM 
and the network (that we are aware of). As it was explained to me by Red 
Hat Support, DLM is extremely efficient, being able to master/distribute 
thousands of locks per second between nodes.

> 
> As I said, lots of questions here :)
> 

and some answers. Wish I had answers to all your questions.

greetings
Riaan
-------------- next part --------------
A non-text attachment was scrubbed...
Name: riaan.vcf
Type: text/x-vcard
Size: 324 bytes
Desc: not available
URL: <http://listman.redhat.com/archives/linux-cluster/attachments/20061025/5fc62eff/attachment.vcf>


More information about the Linux-cluster mailing list