How big should an MGS filesystem be? I would imagine it would be tiny, but I haven''t seen it documented anywhere... Also, is it reasonable to do IP failover for the MGS (assuming shared storage), until MGS failover is implemented by the client code? Thanks, Brent
Brent A Nelson wrote:> How big should an MGS filesystem be? I would imagine it would be tiny, > but I haven''t seen it documented anywhere...Yes, tiny. Just big enough to hold all the config files you would ever need, on the order of 15kb per target (OST or MDT) . You can conceivable pump these up by continually changing parameters, but running a --writeconf on the MDT would reset the logs to their minimum sizes.> > Also, is it reasonable to do IP failover for the MGS (assuming shared > storage), until MGS failover is implemented by the client code? >Yes, assuming that the clients don''t realize the MGS has been lost. If they do, they never try to reconnect to the MGS. In the bigger picture, there is no state associated with an MGS; the only reason there exists an MGC after the initial startup of a client or server is in order to allow the MGS to signal a "new configuration" event, causing the clients/servers to try to read any updates to their configuration from the MGS. For anyone that has lost the MGS connection, they won''t get these updates (until they are restarted), but will continue to run happily with the original configuration.> Thanks, > > Brent > > _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss@clusterfs.com > https://mail.clusterfs.com/mailman/listinfo/lustre-discuss >
Hi, Indeed the MGS storage size can be very small, a few 100MB would be ample. We will fix the MGS failover issue in the near future. It is a stateful failover, and it is not an issue of keeping the same address. The state that the clients reconstruct on the MGS is the "management lock", which is revoked by the MGS just before it changes the configuration. Last I heard (but it is a long time ago) is that for some subtle reason, Lustre''s socklnd in fact doesn''t even work with failover IP addresses. Of course this would be nice to fix perhaps, but given that most other networks don''t have proxy arping etc it''s never been a hot issue. - Peter -> -----Original Message----- > From: lustre-discuss-bounces@clusterfs.com > [mailto:lustre-discuss-bounces@clusterfs.com] On Behalf Of > Brent A Nelson > Sent: Tuesday, September 19, 2006 3:03 PM > To: lustre-discuss@clusterfs.com > Subject: [Lustre-discuss] MGS size > > How big should an MGS filesystem be? I would imagine it would > be tiny, but I haven''t seen it documented anywhere... > > Also, is it reasonable to do IP failover for the MGS > (assuming shared storage), until MGS failover is implemented > by the client code? > > Thanks, > > Brent > > _______________________________________________ > Lustre-discuss mailing list > Lustre-discuss@clusterfs.com > https://mail.clusterfs.com/mailman/listinfo/lustre-discuss