----- Original Message ----- > From: "Tony Stocker" <tony.stoc...@nasa.gov> > To: "Linux HA Cluster Development List" <linux-ha@lists.linux-ha.org> > Sent: Tuesday, May 20, 2014 8:18:52 AM > Subject: [Linux-HA] Problem with migration, priority, stickiness > > Cluster s/w specs: > Kernel: 2.6.32-431.17.1.el6.x86_64 > OS: CentOS 6.5 > corosync-1.4.1-17.el6_5.1.x86_64 > pacemaker-1.1.10-14.el6_5.3.x86_64 > crmsh-2.0+git46-1.1.x86_64 > > > Attached to this email are two text files, one contains the output of 'crm > configure show' (addresses sanitized) and the other contains the output of > 'crm_simulate -sL' > > Here is the situation, and we've encountered this multiple times now and > I've been unable to solve it: > > * A machine in the cluster fails > * There is a spare node, unused, in the cluster available for > assignment > * The resource group that was on the failed machine, instead of > being put onto the spare, unused node is placed on a node where > another resource group is already running > * The displaced resource group then is launched on the spare, > unused node > > As an example, this morning the following occurred: > > Resource Group NRTMASTER is running on system gpmhac01 > Resource Group NRTPNODE1 is running on system gpmhac02 > Resource Group NRTPNODE2 is running on system gpmhac05 > Resource Group NRTPNODE3 is running on system gpmhac04 > Resource Group NRTPNODE4 is running on system gpmhac03 > > system gpmhac06 is up, available, and unused > > system gpmhac04 fails and powers off > > Resource Group NRTPNODE3 is moved to system gpmhac05 > Resource Group NRTNPODE2 is moved to system gpmhac06 > > > One of the big things that seems to occur here is that while the group > NRTPNODE3 is being launched on gpmhac05, the group NRTPNODE2 is being shut > down simultaneously which is causing race conditions where one start > script is putting a state file in place, while the stop script is erasing > it. This leaves the system in an unuseable state because required files, > parameters, and settings are missing/corrupted. > > Secondly, there is simply no reason to kill a perfectly healthy resource > group, that is operating just fine in order to launch a resource group > whose machine has failed when: > 1. There's a spare node available > 2. The resource groups have equal priority with each other, i.e. > all of the NRTPNODE# resource groups have priority "60" > > > So I really need some help here in getting this setup so that it behaves > the way we *think* it should be doing based on what we understand of the > Pacemaker architecture. Obviously we're missing something since this > resource group "shuffling" occurs when there's a failed system, despite > having an unused, spare node available for immediate use, and has bitten > us several times. The fact that the race condition between startup and > shutdown is also causing the system that is brought up to be useless is > exacerbating the situation immensely. > > Ideally, this is what we want: > > 1. If a system fails, the resources/resource group running on it > are moved to an unused, available system. No other resource > shuffling occurs amongst system occurs. > > 2. If a system fails and there is not an unused, available > system to fail over to, then IF the resource group has a higher > priority than another resource group, the group with the lower > priority is shutdown. Only when that shutdown is complete will > the resource group with the higher priority start its startup of > resources. > > 3. If a system fails and there is not an unused, available > system to fail over to, then IF the resource group has the same > or lower priority to all other resource groups, then it will not > attempt to launch itself on any other node, nor cause any other > resource group to stop or migrate. > > 4. Unless specifically, and manually, ordered to move OR if the > hardware system fails, a resource group should remain on its > current hardware system. It should never be forced to migrate to > a new system because something of equal or lower priority failed > and migrated to a new system. > > 5. We do not need resource groups to fail back to original nodes, > when running we want them to stay running on their current system > until/unless a hardware failure occurs and forces them off the > system, or we manually tell them to move. > > > Can someone please look over our configuration, and the bizzare scores > that I see from the crm_simulate output, and help get me to the point > where I can achieve an HA cluster that doesn't kill healthy resources in > some kind of game of musical chairs when there's an empty chair available. > Can you also tell me why or help me to ensure that a startup doesn't occur > until AFTER a shutdown is completely done? > > > Obviously we're misunderstanding or misapplying something with our > resource stickiness, or resource group priority, or something and we > really need to get this resolved. My job is literally on the line with > this due to these failures to operate in the fashion we expect. So all > help is appreciated. What happens if you set resource-stickiness higher. Say to something like 6000, perhaps that would combat the location constraint scores you have going on. -- Vossel > Thanks > Tony > > -- > This message has been scanned for viruses and > dangerous content by MailScanner, and is > believed to be clean. > > > _______________________________________________ > Linux-HA mailing list > Linux-HA@lists.linux-ha.org > http://lists.linux-ha.org/mailman/listinfo/linux-ha > See also: http://linux-ha.org/ReportingProblems _______________________________________________ Linux-HA mailing list Linux-HA@lists.linux-ha.org http://lists.linux-ha.org/mailman/listinfo/linux-ha See also: http://linux-ha.org/ReportingProblems