[Pacemaker] Issue with ordering

Vladislav Bogdanov bubble at hoster-ok.com
Thu Mar 29 05:40:06 EDT 2012


Hi Florian,

29.03.2012 11:54, Florian Haas wrote:
> On Thu, Mar 29, 2012 at 10:07 AM, Vladislav Bogdanov
> <bubble at hoster-ok.com> wrote:
>> Hi Andrew, all,
>>
>> I'm continuing experiments with lustre on stacked drbd, and see
>> following problem:
> 
> At the risk of going off topic, can you explain *why* you want to do
> this? If you need a distributed, replicated filesystem with
> asynchronous replication capability (the latter presumably for DR),
> why not use a Distributed-Replicated GlusterFS volume with
> geo-replication?

I need fast POSIX fs scalable to tens of petabytes with support for
fallocate() and friends to prevent fragmentation.

I generally agree with Linus about FUSE and userspace filesystems in
general, so that is not an option.

Using any API except what VFS provides via syscalls+glibc is not an
option too because I need access to files from various scripted
languages including shell and directly from a web server written in C.
Having bindings for them all is a real overkill. And it all is in
userspace again.

So I generally have choice of CEPH, Lustre, GPFS and PVFS.

CEPH is still very alpha, so I can't rely on it, although I keep my eye
on it.

GPFS is not an option because it is not free and produced by IBM (can't
say which of these two is more important ;) )

Can't remember why exactly PVFS is a no-go, their site is down right
now. Probably userspace server implementation (although some examples
like nfs server discredit idea of in-kernel servers, I still believe
this is a way to go).

Lustre is widely deployed, predictable and stable. It fully runs in
kernel space. Although Oracle did its best to bury Lustre development,
it is actively developed by whamcloud and company. They have builds for
EL6, so I'm pretty happy with this. Lustre doesn't have any replication
built-in so I need to add it on a lower layer (no rsync, no rsync, no
rsync ;) ). DRBD suits my needs for a simple HA.

But I also need datacenter-level HA, that's why I evaluate stacked DRBD
and tickets with booth.

So, frankly speaking, I decided to go with Lustre not because it is so
cool (it has many-many niceties), but because all others I know do not
suit my needs at all due to various reasons.

Hope this clarifies my point,

Best,
Vladislav




More information about the Pacemaker mailing list