[gpfsug-discuss] Monitor NSD server queue?

Thu Aug 18 15:47:25 BST 2016

Done.

Notification generated at: 18 Aug 2016, 10:46 AM Eastern Time (ET)

ID:                                                93260
Headline:                                    Give sysadmin insight into the inner workings of the NSD server machinery, in particular the queue dynamics
Submitted on:                            18 Aug 2016, 10:46 AM Eastern Time (ET)
Brand:                                          Servers and Systems Software
Product:                                      Spectrum Scale (formerly known as GPFS) - Public RFEs

Link:                                            http://www.ibm.com/developerworks/rfe/execute?use_case=viewRfe&CR_ID=93260

Bob Oesterlin
Sr Storage Engineer, Nuance HPC Grid
507-269-0413

From: <gpfsug-discuss-bounces at spectrumscale.org> on behalf of Yuri L Volobuev <volobuev at us.ibm.com>
Reply-To: gpfsug main discussion list <gpfsug-discuss at spectrumscale.org>
Date: Wednesday, August 17, 2016 at 3:34 PM
To: gpfsug main discussion list <gpfsug-discuss at spectrumscale.org>
Subject: [EXTERNAL] Re: [gpfsug-discuss] Monitor NSD server queue?

Unfortunately, at the moment there's no safe mechanism to show the usage statistics for different NSD queues. "mmfsadm saferdump nsd" as implemented doesn't acquire locks when parsing internal data structures. Now, NSD data structures are fairly static, as much things go, so the risk of following a stale pointer and hitting a segfault isn't particularly significant. I don't think I remember ever seeing mmfsd crash with NSD dump code on the stack. That said, this isn't code that's tested and known to be safe for production use. I haven't seen a case myself where an mmfsd thread gets stuck running this dump command, either, but Bob has. If that condition ever reoccurs, I'd be interested in seeing debug data.

I agree that there's value in giving a sysadmin insight into the inner workings of the NSD server machinery, in particular the queue dynamics. mmdiag should be enhanced to allow this. That'd be a very reasonable (and doable) RFE.

yuri

[nactive hide details for "Oesterlin, Robert" ---08/17/2016 04:45:30 AM---]"Oesterlin, Robert" ---08/17/2016 04:45:30 AM---Hi Aaron You did a perfect job of explaining a situation I've run into time after time - high latenc

From: "Oesterlin, Robert" <Robert.Oesterlin at nuance.com>
To: gpfsug main discussion list <gpfsug-discuss at spectrumscale.org>,
Date: 08/17/2016 04:45 AM
Subject: Re: [gpfsug-discuss] Monitor NSD server queue?
Sent by: gpfsug-discuss-bounces at spectrumscale.org

________________________________

Hi Aaron

You did a perfect job of explaining a situation I've run into time after time - high latency on the disk subsystem causing a backup in the NSD queues. I was doing what you suggested not to do - "mmfsadm saferdump nsd' and looking at the queues. In my case 'mmfsadm saferdump" would usually work or hang, rather than kill mmfsd. But - the hang usually resulted it a tied up thread in mmfsd, so that's no good either.

I wish I had better news - this is the only way I've found to get visibility to these queues. IBM hasn't seen fit to gives us a way to safely look at these. I personally think it's a bug that we can't safely dump these structures, as they give insight as to what's actually going on inside the NSD server.

Yuri, Sven - thoughts?

Bob Oesterlin
Sr Storage Engineer, Nuance HPC Grid

From: <gpfsug-discuss-bounces at spectrumscale.org> on behalf of "Knister, Aaron S. (GSFC-606.2)[COMPUTER SCIENCE CORP]" <aaron.s.knister at nasa.gov>
Reply-To: gpfsug main discussion list <gpfsug-discuss at spectrumscale.org>
Date: Tuesday, August 16, 2016 at 8:46 PM
To: gpfsug main discussion list <gpfsug-discuss at spectrumscale.org>
Subject: [EXTERNAL] [gpfsug-discuss] Monitor NSD server queue?

Hi Everyone,

We ran into a rather interesting situation over the past week. We had a job that was pounding the ever loving crap out of one of our filesystems (called dnb02) doing about 15GB/s of reads. We had other jobs experience a slowdown on a different filesystem (called dnb41) that uses entirely separate backend storage. What I can't figure out is why this other filesystem was affected. I've checked IB bandwidth and congestion, Fibre channel bandwidth and errors, Ethernet bandwidth congestion, looked at the mmpmon nsd_ds counters (including disk request wait time), and checked out the disk iowait values from collectl. I simply can't account for the slowdown on the other filesystem. The only thing I can think of is the high latency on dnb02's NSDs caused the mmfsd NSD queues to back up.

Here's my question-- how can I monitor the state of th NSD queues? I can't find anything in mmdiag. An mmfsadm saferdump NSD shows me the queues and their status. I'm just not sure calling saferdump NSD every 10 seconds to monitor this data is going to end well. I've seen saferdump NSD cause mmfsd to die and that's from a task we only run every 6 hours that calls saferdump NSD.

Any thoughts/ideas here would be great.

Thanks!

-Aaron_______________________________________________
gpfsug-discuss mailing list
gpfsug-discuss at spectrumscale.org
http://gpfsug.org/mailman/listinfo/gpfsug-discuss<https://urldefense.proofpoint.com/v2/url?u=http-3A__gpfsug.org_mailman_listinfo_gpfsug-2Ddiscuss&d=CwMFAg&c=djjh8EKwHtOepW4Bjau0lKhLlu-DxM1dlgP0rrLsOzY&r=LPDewt1Z4o9eKc86MXmhqX-45Cz1yz1ylYELF9olLKU&m=D8iCz340ioiUrtGkAFdKjfgfitPkpOr1nRkkxTRCBn0&s=ncd-C59bavCSUTkgYH1vH4ewOM12Hajhy-KhFtKZK68&e=>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gpfsug.org/pipermail/gpfsug-discuss_gpfsug.org/attachments/20160818/a788aa73/attachment-0002.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.gif
Type: image/gif
Size: 106 bytes
Desc: image001.gif
URL: <http://gpfsug.org/pipermail/gpfsug-discuss_gpfsug.org/attachments/20160818/a788aa73/attachment-0002.gif>