Differences
This shows you the differences between two versions of the page.
| Both sides previous revision Previous revision | |||
|
sanboot:ubuntu_iscsi [2010/08/16 20:51] cinquero |
sanboot:ubuntu_iscsi [2010/08/16 20:55] (current) cinquero |
||
|---|---|---|---|
| Line 228: | Line 228: | ||
| </code> | </code> | ||
| - | === Stabilization/Paranoia === | + | === Stabilization/Paranoia/Warnings === |
| in order to prevent data loss (for example, a corrupted dpkg db), one may do the following stuff: | in order to prevent data loss (for example, a corrupted dpkg db), one may do the following stuff: | ||
| * Enable data journaling. That will cut your maximum write performance in half. | * Enable data journaling. That will cut your maximum write performance in half. | ||
| - | * One may increase node.session.timeo.replacement_timeout in iscsid.conf to avoid escalation of temporary connection problems to upper layers. In *my* situation it (almost) *never* makes sense to tell the ext4 driver about I/O problems because there is no fallback. It would just give me an unclean filesystem (not unmounted properly, like a hard reset). However(!!), an unclean filesystem may be better than a totally screwed one: imagine your iSCSI target host crashes, looses some data that still has not been written to disk, and the client continues to use that target without doing a journal replay first... personally, I try to avoid that situation by never starting the iSCSI target automatically (removed from init scripts, starting it manually after killing the client machines...). IMHO the iSCSI protocol is a bit dumb. The target should detect a crash and refuse session continuation after restart... again, I'm not liable for *any* sort of data loss if you act according to these explanations. You *will* screw your data. Sooner or later. Especially because SAN boot using gpxe seems to not allow using a MaxSessions=1/MaxConnections=1 setting for the target, which would prevent concurrent accesses -- a very basic need for data integrity purposes. | + | * One may increase node.session.timeo.replacement_timeout in iscsid.conf to avoid escalation of temporary connection problems to upper layers. In *my* situation it (almost) *never* makes sense to tell the ext4 driver about I/O problems because there is no fallback. It would just give me an unclean filesystem (not unmounted properly, like a hard reset). However(!!), an unclean filesystem may be better than a totally screwed one: imagine your iSCSI target host crashes, looses some data that still has not been written to disk, and the client continues to use that target without doing a journal replay first... personally, I try to avoid that situation by never starting the iSCSI target automatically (removed from init scripts, starting it manually after killing the client machines...). IMHO the iSCSI protocol is a bit dumb. The target should detect a crash and refuse session continuation after restart... again, I'm not liable for *any* sort of data loss if you act according to these explanations. You *will* screw your data. Sooner or later. |
| + | * Additional warning: SAN boot using gpxe seems to not allow using a MaxSessions=1/MaxConnections=1 setting for the target, which would prevent concurrent accesses -- a very basic need for data integrity purposes. | ||
