 |
 |
On Fri, 22 Oct 2004 11:44:59 +0100 Steve Traylen <[log in to unmask]> wrote: Hi Steve I'm planning to create a LVM with one massive partition for our SE. With reguard to XFS, we have another RAID system connected to our group backup server. We have configured it with LVM and jfs (from IBM), the only problem we have discovered with jfs is that there is no quota support yet, for this particular application this is not a problem as we don't want to quota that particular partition. However, I did look at XFS as an alternative to jfs but I discovered that it isn't supported by Redhat Enterprise 3.2 which is the operating system we are using. I am wondering if this is true of Scientific Linux as well i.e. not supporting XFS. Mark. > Hi, > > There are mails coming in from various directions asking about > how to currently make use of many partitions of 1 or 2 TB > within LCG on an SE. > > The current situation is this in that per ClassicSE you can only > have one disk partition per VO per SE. > > There are few ways to consider to solve this, all of which have > problems and as a result leave you fairly stuck. > > 1) On the same host you can create a RAID5 with LCM or something > and put XFS on it. XFS gives you a maximum partition of at least > a PB though I'm not sure of the exact size. > > However we are easily able to break and LVM system at RAL > with out to much difficulty. Also XFS in RH7.3 is best avoided. > > 2) Use some kind on SAN technology, eg GFS, iSCSI to give to huge > massive file system. Again you will need to use XFS on top. > > 3) Wait for dcache as a disk pool manager. > We are having a bit more success recently with this here and > I am feeling more confident about it. > We would hope to have a 8 TB (4 partitions of 2TB on 2 hosts) > some time next week. We've done it once now and hope to now be > able to deploy it again in a more automated way. > For those considering this option and wanting to plan for it I > would recommend that you will need a front end DCache HEAD node > for your storage. One of the GridPP provides boxes is just > perfect. Putting a mirrored pair or something on the head node > may be a very good idea since loosing the head node and the > database on it results in the loss of ALL of your data across > systems. > > We will provide some more information, pointers to documentation > etc if we have success next week. > > 4) Internal releases of the edg-se claim to be able to help but is > not > an option for deployment today. > > Steve > > > > -- > Steve Traylen > [log in to unmask] > http://www.gridpp.ac.uk/