[Linux-cluster] GFS Performance advise

[Tomer Okavi]

Thanks Leonardo, you cleared things up for me.

Tom Ok.

On 7/26/06, Leonardo Rodrigues de Mello <Leonardo.Mello at planejamento.gov.br>
wrote:
>
> Other things I have forgot in the last message:
>
> 1- You dont need to use the SAN to have one ACTIVE/PASSIVE cluster, with
> data replication between the servers. Check out DRBD 0.7 + Heartbeat or
> RH-Cluster-Suite. This will be able to do the job. without the cost of one
> external storage device.
> a good point to start is:
>
> 2- RedHat are developing one Cluster Raid aproach that will be better than
> drbd because it will be possible to create a distributed raid to split the
> storage between the servers. I dont know how the development of this draid
> is going or when it will be made stable. Does anyone from RH can tell
> something about this topic ?
>
>
> Some links for you:
> a - Official documentation
>
> http://www.drbd.org/documentation.html
>
> b - drbd instalation
> Portuguese:
>
> http://guialivre.governoeletronico.gov.br/mediawiki/index.php/DocumentacaoTecnologiasDRBD
> English:
> http://www.linux-ha.org/DRBD/HowTo/Install
> http://linux-ha.org/DRBD/QuickStart07
>
>
> c - drbd + heartbeat integration
> http://www.slackworks.com/~dkrovich/DRBD/heartbeat.html
> http://www.linux-ha.org/GettingStarted/DRBD
>
> best regards
> Leonardo Rodrigues de Mello
> -----Mensagem original-----
> De:     Leonardo Rodrigues de Mello em nome de Leonardo Rodrigues de Mello
> Enviada:        qua 26/7/2006 09:36
> Para:   linux clustering
> Cc:
> Assunto:        RES: [Linux-cluster] GFS Performance advise
>
> Gfs is only necessary if you have two or more machines that access
> (READ+WRITE) the filesystem at the same time. GFS will create and manage a
> global lock of the filesystem, and other things to make shure the filesystem
> can be shared among the cluster nodes without filesystem corruption.
>
> beside that fact, if you have active/passive you can use ext3 without any
> problem. you can use gfs no_lock too. if you are having problems with gfs
> no_lock maybe because something is misconfigured in your setup. You CANT use
> gfs no_lock the same way you use gfs with dlm or gulm because if you do that
> you can get a filesystem corruption... i dont know if gfs permit one
> configuration like that.
>
> best regards
> Leonardo Rodrigues de Mello
>
>
> -----Mensagem original-----
> De:     linux-cluster-bounces at redhat.com em nome de Tomer Okavi
> Enviada:        qua 26/7/2006 03:17
> Para:   linux-cluster at redhat.com
> Cc:
> Assunto:        [Linux-cluster] GFS Performance advise
>
> I've a samba file server cluster (Active\Passive) with 2 cluster nodes on
> Cent OS 4.3
> both nodes are connected to a shared storage through Fiber switch+HBA
> the shared storage holds the file system that samba shares to the windows
> machines.
> only one cluster node mounts the file system (the active one)
> currently I'm using ext3 as the file system on the shared storage because
> I've experienced slow response time and locking problems from the samba
> service.
> I've tried formatting the shared file system with GFS disabling locks
> (lock_nolock), tried mounting the file system with
> lockproto=lock_nolock,localchaching,localflocks
> with no success, samba still complains about oplocks breakes and the
> windows
> system connecting to the shares experience slow performance from samba.
> the samba file system exports the file system to 3 IIS servers through unc
> path's
> it's dealing with lots (1,000,000) of small (under 250KB) files.
> when using ext3 as the file system for the samba shares i have no problem.
>
>
> 1. should i use GFS for the file system?, to avoid file system corruption
> in
> case one cluster node crash or is ext3 is a good enough solution?
> 2. why when using GFS with lockproto=lock_nolock,localchaching,localflocks
> i
> still see "glock nq calls" and  "lm_lock calls" in gfs_tool counters
>
> my main goal is to achieve maximum samba performance with the lowest
> chance
> for file system corruption in case of a failover or crashed cluster node.
>
> thanks
>
> Tom Ok.
>
>
>
>
>
> --
> Linux-cluster mailing list
> Linux-cluster at redhat.com
> https://www.redhat.com/mailman/listinfo/linux-cluster
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/linux-cluster/attachments/20060727/41cabd9e/attachment.htm>