Another possibility is to try increasing the timeouts. We used to have problems with this all of the time on clusters with thousands of nodes, but now we run with the following settings increased from their [defaults]…
sqtBusyThreadTimeout [10] = 120 sqtCommandRetryDelay [60] = 120 sqtCommandTimeout [300] = 500 These are in the category of undocumented configurables, so you may wish to accompany this with a PMR. And you’ll need to know the secret handshake that follows this… mmchconfig: Attention: Unknown attribute specified: sqtBusyThreadTimeout. Press the ENTER key to continue. -Paul From: [email protected] <[email protected]> On Behalf Of Sven Oehme Sent: Thursday, February 20, 2020 17:29 To: gpfsug main discussion list <[email protected]> Subject: Re: [gpfsug-discuss] Unkillable snapshots This message was sent by an external party. Filesystem quiesce failed has nothing to do with open files. What it means is that the filesystem couldn’t flush dirty data and metadata within a defined time to take a snapshot. This can be caused by to high maxfilestocache or pagepool settings. To give you an simplified example (its more complex than that, but good enough to make the point) - assume you have 100 nodes, each has 16 GB pagepool and your storage system can write data out at 10 GB/sec, it will take 160 seconds to flush all data data (assuming you did normal buffered I/O. If i remember correct (talking out of memory here) the default timeout is 60 seconds, given that you can’t write that fast it will always timeout under this scenario. There is one case where this can also happen which is a client is connected badly (flaky network or slow connection) and even your storage system is fast enough the node is too slow that it can’t de-stage within that time while everybody else can and the storage is not the bottleneck. Other than that only solutions are to a) buy faster storage or b) reduce pagepool and maxfilestocache which will reduce overall performance of the system. Sven Sent from my iPad On Feb 20, 2020, at 5:14 PM, Nathan Falk <[email protected]<mailto:[email protected]>> wrote: Good point, Simon. Yes, it is a "file system quiesce" not a "fileset quiesce" so it is certainly possible that mmfsd is unable to quiesce because there are processes keeping files open in another fileset. Nate Falk IBM Spectrum Scale Level 2 Support Software Defined Infrastructure, IBM Systems From: Simon Thompson <[email protected]<mailto:[email protected]>> To: gpfsug main discussion list <[email protected]<mailto:[email protected]>> Date: 02/20/2020 04:39 PM Subject: [EXTERNAL] Re: [gpfsug-discuss] Unkillable snapshots Sent by: [email protected]<mailto:[email protected]> ________________________________ Hi Nate, So we're trying to clean up snapshots from the GUI ... we've found that if it fails to delete one night for whatever reason, it then doesn't go back another day and clean up 😊 But yes, essentially running this by hand to clean up. What I have found is that lsof hangs on some of the "suspect" nodes. But if I strace it, its hanging on a process which is using a different fileset. For example, the file-set we can't delete is: rds-projects-b which is mounted as /rds/projects/b But on some suspect nodes, strace lsof /rds, that hangs at a process which has open files in: /rds/projects/g which is a different file-set. What I'm wondering if its these hanging processes in the "g" fileset which is killing us rather than something in the "b" fileset. Looking at the "g" processes, they look like a weather model and look to be dumping a lot of files in a shared directory, so I wonder if the mmfsd process is busy servicing that and so whilst its not got "b" locks, its just too slow to respond? Does that sound plausible? Thanks Simon ________________________________ From: [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>> on behalf of [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>> Sent: 20 February 2020 21:26:39 To: gpfsug main discussion list Subject: Re: [gpfsug-discuss] Unkillable snapshots Hello Simon, Sadly, that "1036" is not a node ID, but just a counter. These are tricky to troubleshoot. Usually, by the time you realize it's happening and try to collect some data, things have already timed out. Since this mmdelsnapshot isn't something that's on a schedule from cron or the GUI and is a command you are running, you could try some heavy-handed data collection. You suspect a particular fileset already, so maybe have a 'mmdsh -N all lsof /path/to/fileset' ready to go in one window, and the 'mmdelsnapshot' ready to go in another window? When the mmdelsnapshot times out, you can find the nodes it was waiting on in the file system manager mmfs.log.latest and see what matches up with the open files identified by lsof. It sounds like you already know this, but the <c0n42> type of internal node names in the log messages can be translated with 'mmfsadm dump tscomm' or also plain old 'mmdiag --network'. Thanks, Nate Falk IBM Spectrum Scale Level 2 Support Software Defined Infrastructure, IBM Systems From: Simon Thompson <[email protected]<mailto:[email protected]>> To: gpfsug main discussion list <[email protected]<mailto:[email protected]>> Date: 02/20/2020 03:14 PM Subject: [EXTERNAL] Re: [gpfsug-discuss] Unkillable snapshots Sent by: [email protected]<mailto:[email protected]> ________________________________ Hmm ... mmdiag --tokenmgr shows: Server stats: requests 195417431 ServerSideRevokes 120140 nTokens 2146923 nranges 4124507 designated mnode appointed 55481 mnode thrashing detected 1036 So how do I convert "1036" to a node? Simon ________________________________ From: [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>> on behalf of Simon Thompson <[email protected]<mailto:[email protected]>> Sent: 20 February 2020 19:45:02 To: gpfsug main discussion list Subject: [gpfsug-discuss] Unkillable snapshots Hi, We have a snapshot which is stuck in the state "DeleteRequired". When deleting, it goes through the motions but eventually gives up with: Unable to quiesce all nodes; some processes are busy or holding required resources. mmdelsnapshot: Command failed. Examine previous error messages to determine cause. And in the mmfslog on the FS manager there are a bunch of retries and "failure to quesce" on nodes. However in each retry its never the same set of nodes. I suspect we have one HPC job somewhere killing us. What's interesting is that we can delete other snapshots OK, it appears to be one particular fileset. My old goto "mmfsadm dump tscomm" isn't showing any particular node, and waiters around just tend to point to the FS manager node. So ... any suggestions? I'm assuming its some workload holding a lock open or some such, but tracking it down is proving elusive! Generally the FS is also "lumpy" ... at times it feels like a wifi connection on a train using a terminal, I guess its all related though. Thanks Simon _______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org http://gpfsug.org/mailman/listinfo/gpfsug-discuss _______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org http://gpfsug.org/mailman/listinfo/gpfsug-discuss _______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org http://gpfsug.org/mailman/listinfo/gpfsug-discuss
_______________________________________________ gpfsug-discuss mailing list gpfsug-discuss at spectrumscale.org http://gpfsug.org/mailman/listinfo/gpfsug-discuss
