On Thu, 13 Sep 2007, Frankie Montenegro wrote:
Well, i evaluated gfs over gnbd and bonded interfaces a few months ago and
was quite happy with the performance. But it was crashing quite often
and i didn't need the complexity of quorum and fencing issues when looking
for just a cluster file system and not high availability features. I guess
there is no way around complex fencing when every node accesses the block
device through the cluster file system. Its the same with ocfs2 too, but
fencing is much simpler with ocfs2 and less number of daemons to mess
with compared to gfs, where quite a few daemons need to come up in a
particular order.
I have been using Lustre for almost 2 yrs now and quite happy
with it. I posted a few problems about gnbd errors before crashes, didn't
get much help. So went back to Lustre and I am still satisfied with the
choice. I was evaluating the cluster file systems for the two clusters i
finished in the past 2 months, the first one with one storage server and 6
TB of storage and second one with >65TB of available storage space with
multiple storage servers, bonding, parallel i/o and all bells and whistles. Both
run Lustre. The new version is quite easy to maintain. Since i was already
comfortable with Lustre, my opinions could be biased.
But to be fair, the throughput of gfs over gnbd was very good. But
lack of parallel i/o and having to deal with quorum issues when all you
need is a cluster file system, make me go back to Lustre. Hope this
helps!!
Regards
Balagopal
> Will check that out. Thank you guys. My stuff has arrived, I can't wait for
> the weekend to put it together.
>
> BTW, I wondered what FS would be the best
> option for the good throughput as well as sharing of the disk space. We have
> NFS at school on a 12node dual opteron cluster, and, based on my experience
> with it, I'll try to stay as far away from it as possible. GFS or PVFS2 or...?
> Thanks again,
> Frankie
>
> On 9/13/07, Nate Carlson <natecars@xxxxxxxxxxxxxxx> wrote:
> > On Fri, 7 Sep 2007, Balagopal Pillai wrote:
> > > For 802.3ad bonding mode, the switch needs to support lacp. Static
> > > trunking feature on the switch is not enough. In your case with no
> > > switch support, mode 6 or adaptive load balancing is a good option.
> > > Round robin is the only mode that will give you more than an interface
> > > worth of throughput on a single connection. But that needs some switch
> > > support. (like cisco etherchannel for example) Also there is additional
> > > overhead due to out of the order packets. The other modes will give
> > > better aggregate throughput.
> >
> > Well, if you use crossover cables, you can do it without switch support..
> > ;)
> >
> > ------------------------------------------------------------------------
> > | nate carlson | natecars@xxxxxxxxxxxxxxx | http://www.natecarlson.com |
> > | depriving some poor village of its idiot since 1981 |
> > ------------------------------------------------------------------------
> >
> > --
> > Linux-cluster mailing list
> > Linux-cluster@xxxxxxxxxx
> > https://www.redhat.com/mailman/listinfo/linux-cluster
> >
>
--
Linux-cluster mailing list
Linux-cluster@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/linux-cluster
