At 05:08 AM 1/12/2009, Stas Oskin wrote: > From Lustre wiki > (<http://wiki.lustre.org/index.php?title=Lustre_FAQ>http://wiki.lustre.org/index.php?title=Lustre_FAQ) > > >Are fibrechannel switches necessary? How does HA shared storage work? > > > >Typically, fibrechannel switches are not necessary. Multi-port >shared storage for failover is normally configured to be shared >between two server nodes on a FC-AL. Shared SCSI and future shared >SATA devices will also work. > >Backend storage is expected to be cache-coherent between multiple >channels reaching the devices. Servers in an OSS failover pair are >normally both active in the file system, and can be configured to >take over partitions for each other in the case of a failure. MDS >failover pairs can also both be active, but only if they serve >multiple separate file systems. > > >As far a I understand, Lustre is designed with the approach most >cluster file systems (except GlusterFS of course :) ), meaning you >have master servers that responsible for storage and retrieval of >the data, and storage nodes, which do actualy storage. unless something has changed drastically since the last time I read up on lustre, it requires shared storage in some capacity. Whether they're connected via a SCSI ring or FC-AL or iSCSI or whatever other mechanism, as I understand it, the machine needs to view the physical disks as devices on the machine, and mirrors across the block devices. This is similar to ocfs2 which I think is superior to lustre, but still kind of difficult to configure. The advantages gluster has over all the other cluster filesystems out there are: 1) you can use commodity hardware without having to build additional infrastructure (you don't need a SAN, just the same computers connected to the same network will do the trick) 2) replica's (ha/afr, whatever) needn't be physically connected (or connected by anything but a network connection. While it might be a tad slow, it will work exactly the same over a wide area as within a data center. While number 2 could be addressed with a remote block device (drbd) when used with ocfs2 or probably lustre, they require more complex administration.
