>>> Valentin Vidic <[email protected]> schrieb am 03.04.2019 um 09:26 in Nachricht <[email protected]>: > On Wed, Apr 03, 2019 at 09:13:58AM +0200, Ulrich Windl wrote: >> I'm surprised: Once sbd writes the fence command, it usually takes >> less than 3 seconds until the victim is dead. If you power off a >> server, the PDU still may have one or two seconds "power reserve", so >> the host may not be down immediately. Besides of that power‑cycles are >> additional stress for the hardware... >> >> So maybe you want to explain why and how much faster IPMI and PDU fencing > are. > > SBD is slow for me too. Since it doesn't have a way to confirm the kill > it needs to wait for various timeouts and these can be quite high. For > example the IBM storage timeouts require this setup: > > Timeout (watchdog) : 130 > Timeout (msgwait) : 270
Reminds me of a 30 minute SCSI timeout for the rewind command to a magnetic tape... SCSI still has (AFAIK) 60s disk timeouts, but actually if a disk's response time exceeds 3 seconds on reads, it's time to replace it. I don't know what the IBM storage system does behind the scenes, but the timeout seems excessively slow. > > On the same cluster IPMI fence executes in a second or two, but requires > network connectivity. > > ‑‑ > Valentin > _______________________________________________ > Manage your subscription: > https://lists.clusterlabs.org/mailman/listinfo/users > > ClusterLabs home: https://www.clusterlabs.org/ _______________________________________________ Manage your subscription: https://lists.clusterlabs.org/mailman/listinfo/users ClusterLabs home: https://www.clusterlabs.org/
