If the public write ring between the two halves of a resilient multi node system is broken then the system will crash when the next public write is attempted. After this crash each half of the system will automatically restart load [depending on IPL defaults] on simplex filestore.

For systems that have implemented multi-plexed system filestore then a copy of the catalogue will now exist in both halves of the system. Once the individual systems have restart loaded then VME increments the timestamp in the partition containing the Catalogue by 2 seconds. Since both systems have reloaded then this will occur on each plex of the catalogue - thus the plexes will still have the same timestamp.

In the Catalogue for system A the Catalogue plex A will be marked as active, but plex B will be marked as broken. In system B will be the reverse : plex A will be marked as broken and plex B will be active.

Assuming that system A is then run as the "prime" system then a potential problem may arise once the link between the two systems is re-established.

When performing this sort of activity the following must be performed. Identify which half of the system was considered to be the prime system. A full load must then be performed to reload this prime system.

This will ensure the timestamp on this plex of the catalogue will be updated with the present time, which means the timestamp will be different from plex B. When system B is rejoined to system A then plex B will be detected as being different to plex A and plex B will be reinstated from plex A.

It is also recommended that sites install a check in the SITE_RESTART_SYSTEM hook to ensure that the load is as expected. If it is not, prevent any work from starting and inform the operator of the problem.

Fujitsu Services are currently reviewing VME plexing and the way it affects VME loading. Information on any changes in this area will be published as soon as it is available.