[Gluster-devel] cluster/stripe

[Alexey Filin]

Hi,

gluster at sv.gnu.org/glusterfs--mainline--2.5--patch-485
fuse-2.7.0-glfs3
Linux 2.6.9-55.0.2.EL.cern (Scientific Linux CERN SLC release
4.5(Beryllium)), i386

4 HPC cluster work nodes, each node has two Gigabit interfaces for two LANs
(Data Acquisition System LAN and SAN).

server.vol and client.vol were made with example in
http://gluster.org/docs/index.php/GlusterFS_Translators but with alu
scheduler:

brick->posix-locks->io-thr->wb->ra->server
((client1+client2)->stripe1)+((client3+client4)->stripe2)->afr->unify->iot->wb->ra->ioc

unify is supposed to connect another 4 nodes after tests

copying from local FS to GlusterFS and inversely on client1 works fine,
performance nearly native (as for local to local)
back-end FS ext3 (get/setfattr don't work)=> afr works fine, stripe doesn't
work at all
back-end FS xfs (get/setfattr work fine) => afr works fine, stripe doesn't
work at all

changed client.vol to: (client1+client2)->stripe1->iot->wb->ra->ioc =>
stripe doesn't work

log files don't contain anything interesting,
1) How to make cluster/stripe work?

http://gluster.org/docs/index.php/GlusterFS_FAQ
"...if one uses 'cluster/afr' translator with 'cluster/stripe' then
GlusterFS can provide high availability."
2) Is HA provided for stripe+afr only or for afr alone too?

I suppose to use cluster work nodes with local hard disks as a distributed
on-line and off-line storage for raw data acquired on our experimental setup
(tape back-up is provided of course). It's supposed to keep 10-20 terabytes
of raw data in total (the cluster is supposed to upgrade in future).
3) Has it sense to use cluster/stripe (HA is very desirable) in the case?

Thanks for answers in advance.

Alexey Filin.
Experiment OKA, Institute for High Energy Physics, Protvino, Russia

PS Also I tortured GlusterFS with direct file manipulation (through back-end
FS), results are good for me