[gpfsug-discuss] Same file opened by many nodes / processes
Peter Childs
p.childs at qmul.ac.uk
Mon Jul 23 21:06:14 BST 2018
---- Frederick Stock wrote ----
> Have you considered keeping the 1G network for daemon traffic and moving the data traffic to another network?
Considered, but never really understood the logic or value of building a second network, nor seen a good argument for the additional cost and work setting it up. While I've heard it lots of times, that the network is key to good gpfs performance. I've actually always found that it can be lots of other things too and your usally best keeping and open view and checking everything.
This issue disappeared on Friday when the file system manager locked up entirely, and we failed it over to the other one and restarted gpfs. It's been fine all weekend, and currently it's looking to be a failed gpfs daemon on the manager node that was causing all the bad io.
If I'd know that I'd have restarted gpfs on that node earlier...
>
> Given the description of your configuration with only 2 manager nodes handling mmbackup and other tasks my guess is that is where the problem lies regarding performance when mmbackup is running with the many nodes accessing a single file. You said the fs managers were on hardware, does that mean other nodes in this cluster are VMs of some kind?
>
> You stated that your NSD servers were under powered. Did you address that problem in any way, that is adding memory/CPUs, or did you just move other GPFS activity off of those nodes?
>
Our nsd servers are virtual everything else on the cluster is real. It's a gridscaler gs7k. Hence why it's difficult to throw more power at the issue.
We are looking at upgrading to 5.0.1<tel:5.0.1>, within the next few months as we're in the progress of adding a new ssd based scratch filesystem to the cluster.
Hopefully this will help resolve some of our issues.
Peter Childs.
> Fred
> __________________________________________________
> Fred Stock | IBM Pittsburgh Lab | 720-430-8821<tel:720-430-8821>
> stockf at us.ibm.com<mailto:stockf at us.ibm.com>
>
>
>
> From: Peter Childs <p.childs at qmul.ac.uk<mailto:p.childs at qmul.ac.uk>>
> To: "gpfsug-discuss at spectrumscale.org<mailto:gpfsug-discuss at spectrumscale.org>" <gpfsug-discuss at spectrumscale.org<mailto:gpfsug-discuss at spectrumscale.org>>
> Date: 07/23/2018 07:06 AM
> Subject: Re: [gpfsug-discuss] Same file opened by many nodes / processes
> Sent by: gpfsug-discuss-bounces at spectrumscale.org<mailto:gpfsug-discuss-bounces at spectrumscale.org>
>
>
>
> On Mon, 2018-07-23<tel:2018-07-23> at 22:13 +1200<tel:+1200>, José Filipe Higino wrote:
> I think the network problems need to be cleared first. Then I would investigate further.
>
> Buf if that is not a trivial path...
> Are you able to understand from the mmfslog what happens when the tipping point occurs?
>
> mmfslog thats not a term I've come accross before, if you mean /var/adm/ras/mmfs.log.latest<http://mmfs.log.latest> then I'm already there is not a lot there, In other words no expulsions or errors just a very slow filesystem, We've not seen any significantly long waiters either (mmdiag --waiters) so as far as I can see its just behaving like a very very busy filesystem.
>
> We've already had IBM looking at the snaps due to the rather slow mmbackup process, all I've had back is to try increase -a ie the number of sort threads which has speed it up to a certain extent, But once again I think we're looking at the results of the issue not the cause.
>
>
> In my view, when troubleshooting is not easy, the usual methods work/help to find the next step:
> - Narrow the window of troubleshooting (by discarding "for now" events that did not happen within the same timeframe)
> - Use "as precise" as possible, timebased events to read the reaction of the cluster (via log or others) and make assumptions about other observed situations.
> - If possible and when the problem is happening, run some traces, gpfs.snap<http://gpfs.snap> and ask for support via PMR.
>
> Also,
>
> What is version of GPFS?
>
> 4.2.3-8<tel:4.2.3-8>
>
> How many quorum nodes?
>
> 4 Quorum nodes with tie breaker disks, however these are not the file system manager nodes as to fix a previous problem (with our nsd servers not being powerful enough) our fsmanager nodes are on hardware, We have two file system manager nodes (Which do token management, quota management etc) they also run the mmbackup.
>
> How many filesystems?
>
> 1, although we do have a second that is accessed via multi-cluster from our older GPFS setup, (thats running 4.2.3-6<tel:4.2.3-6> currently)
>
> Is the management network the same as the daemon network?
>
> Yes. the management network and the daemon network are the same network.
>
> Thanks in advance
>
> Peter Childs
>
>
>
> On Mon, 23 Jul 2018 at 20:37, Peter Childs <p.childs at qmul.ac.uk<mailto:p.childs at qmul.ac.uk>> wrote:
> On Mon, 2018-07-23<tel:2018-07-23> at 00:51 +1200<tel:+1200>, José Filipe Higino wrote:
>
> Hi there,
>
> Have you been able to create a test case (replicate the problem)? Can you tell us a bit more about the setup?
>
> Not really, It feels like a perfect storm, any one of the tasks running on its own would be fine, Its the shear load, our mmpmon data says the storage has been flat lining when it occurs.
>
> Its a reasonably standard (small) HPC cluster, with a very mixed work load, hence while we can usually find "bad" jobs from the point of view of io on this occasion we can see a few large array jobs all accessing the same file, the cluster runs fine until we get to a certain point and one more will tip the balance. We've been attempting to limit the problem by adding limits to the number of jobs in an array that can run at once. But that feels like fire fighting.
>
>
> Are you using GPFS API over any administrative commands? Any problems with the network (being that Ethernet or IB)?
>
> We're not as using the GPFS API, never got it working, which is a shame, I've never managed to figure out the setup, although it is on my to do list.
>
> Network wise, We've just removed a great deal of noise from arp requests by increasing the arp cache size on the nodes. Its a mixed 1GBit/10GBit network currently, we're currently looking at removing all the 1GBit nodes within the next few months and adding some new faster kit. The Storage is attached at 40GBit but it does not look to want to run much above 5Gbit I suspect due to Ethernet back off due to the mixed speeds.
>
> While we do have some IB we don't currently run our storage over it.
>
> Thanks in advance
>
> Peter Childs
>
>
>
>
>
> Sorry if I am un-announced here for the first time. But I would like to help if I can.
>
> Jose Higino,
> from NIWA
> New Zealand
>
> Cheers
>
> On Sun, 22 Jul 2018 at 23:26, Peter Childs <p.childs at qmul.ac.uk<mailto:p.childs at qmul.ac.uk>> wrote:
> Yes, we run mmbackup, using a snapshot.
>
> The scan usally takes an hour, but for the last week has been taking many hours (i saw it take 12 last Tuesday)
>
> It's speeded up again now back to its normal hour, but the high io jobs accessing the same file from many nodes also look to have come to an end for the time being.
>
> I was trying to figure out howto control the bad io using mmchqos, to prioritise certain nodes over others but had not worked out if that was possible yet.
>
> We've only previously seen this problem when we had some bad disks in our storage, which we replaced, I've checked and I can't see that issue currently.
>
> Thanks for the help.
>
>
>
> Peter Childs
> Research Storage
> ITS Research and Teaching Support
> Queen Mary, University of London
>
> ---- Yaron Daniel wrote ----
>
> Hi
>
> Do u run mmbackup on snapshot , which is read only ?
>
>
> Regards
>
>
>
> Yaron Daniel 94 Em Ha'Moshavot Rd
>
> Storage Architect – IL Lab Services (Storage) Petach Tiqva, 49527<tel:49527> IBM Global Markets, Systems HW Sales Israel Phone: +972-3-916-5672<tel:+972-3-916-5672> Fax: +972-3-916-5672<tel:+972-3-916-5672> Mobile: +972-52-8395593<tel:+972-52-8395593> e-mail: yard at il.ibm.com<mailto:yard at il.ibm.com> IBM Israel
>
>
>
>
>
>
> From: Peter Childs <p.childs at qmul.ac.uk<mailto:p.childs at qmul.ac.uk>>
> To: "gpfsug-discuss at spectrumscale.org<mailto:gpfsug-discuss at spectrumscale.org>" <gpfsug-discuss at spectrumscale.org<mailto:gpfsug-discuss at spectrumscale.org>>
> Date: 07/10/2018 05:51 PM
> Subject: [gpfsug-discuss] Same file opened by many nodes / processes
> Sent by: gpfsug-discuss-bounces at spectrumscale.org<mailto:gpfsug-discuss-bounces at spectrumscale.org>
>
>
>
>
> We have an situation where the same file is being read by around 5000
> "jobs" this is an array job in uge with a tc set, so the file in
> question is being opened by about 100 processes/jobs at the same time.
>
> Its a ~200GB file so copying the file locally first is not an easy
> answer, and these jobs are causing issues with mmbackup scanning the
> file system, in that the scan is taking 3 hours instead of the normal
> 40-60<tel:40-60> minutes.
>
> This is read only access to the file, I don't know the specifics about
> the job.
>
> It looks like the metanode is moving around a fair amount (given what I
> can see from mmfsadm saferdump file)
>
> I'm wondering if we there is anything we can do to improve things or
> that can be tuned within GPFS, I'm don't think we have an issue with
> token management, but would increasing maxFileToCache on our token
> manager node help say?
>
> Is there anything else I should look at, to try and attempt to allow
> GPFS to share this file better.
>
> Thanks in advance
>
> Peter Childs
>
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org<http://spectrumscale.org>
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss
>
> --
> Peter Childs
> ITS Research Storage
> Queen Mary, University of London
>
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org<http://spectrumscale.org>
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org<http://spectrumscale.org>
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss
>
> --
> Peter Childs
> ITS Research Storage
> Queen Mary, University of London
> _______________________________________________
> gpfsug-discuss mailing list
> gpfsug-discuss at spectrumscale.org<http://spectrumscale.org>
> http://gpfsug.org/mailman/listinfo/gpfsug-discuss
Have you considered keeping the 1G network for daemon traffic and moving the data traffic to another network?
Given the description of your configuration with only 2 manager nodes handling mmbackup and other tasks my guess is that is where the problem lies regarding performance when mmbackup is running with the many nodes accessing a single file. You said the fs managers were on hardware, does that mean other nodes in this cluster are VMs of some kind?
You stated that your NSD servers were under powered. Did you address that problem in any way, that is adding memory/CPUs, or did you just move other GPFS activity off of those nodes?
Fred
__________________________________________________
Fred Stock | IBM Pittsburgh Lab | 720-430-8821
stockf at us.ibm.com
From: Peter Childs <p.childs at qmul.ac.uk>
To: "gpfsug-discuss at spectrumscale.org" <gpfsug-discuss at spectrumscale.org>
Date: 07/23/2018 07:06 AM
Subject: Re: [gpfsug-discuss] Same file opened by many nodes / processes
Sent by: gpfsug-discuss-bounces at spectrumscale.org
________________________________
On Mon, 2018-07-23 at 22:13 +1200, José Filipe Higino wrote:
I think the network problems need to be cleared first. Then I would investigate further.
Buf if that is not a trivial path...
Are you able to understand from the mmfslog what happens when the tipping point occurs?
mmfslog thats not a term I've come accross before, if you mean /var/adm/ras/mmfs.log.latest then I'm already there is not a lot there, In other words no expulsions or errors just a very slow filesystem, We've not seen any significantly long waiters either (mmdiag --waiters) so as far as I can see its just behaving like a very very busy filesystem.
We've already had IBM looking at the snaps due to the rather slow mmbackup process, all I've had back is to try increase -a ie the number of sort threads which has speed it up to a certain extent, But once again I think we're looking at the results of the issue not the cause.
In my view, when troubleshooting is not easy, the usual methods work/help to find the next step:
- Narrow the window of troubleshooting (by discarding "for now" events that did not happen within the same timeframe)
- Use "as precise" as possible, timebased events to read the reaction of the cluster (via log or others) and make assumptions about other observed situations.
- If possible and when the problem is happening, run some traces, gpfs.snap and ask for support via PMR.
Also,
What is version of GPFS?
4.2.3-8
How many quorum nodes?
4 Quorum nodes with tie breaker disks, however these are not the file system manager nodes as to fix a previous problem (with our nsd servers not being powerful enough) our fsmanager nodes are on hardware, We have two file system manager nodes (Which do token management, quota management etc) they also run the mmbackup.
How many filesystems?
1, although we do have a second that is accessed via multi-cluster from our older GPFS setup, (thats running 4.2.3-6 currently)
Is the management network the same as the daemon network?
Yes. the management network and the daemon network are the same network.
Thanks in advance
Peter Childs
On Mon, 23 Jul 2018 at 20:37, Peter Childs <p.childs at qmul.ac.uk<mailto:p.childs at qmul.ac.uk>> wrote:
On Mon, 2018-07-23 at 00:51 +1200, José Filipe Higino wrote:
Hi there,
Have you been able to create a test case (replicate the problem)? Can you tell us a bit more about the setup?
Not really, It feels like a perfect storm, any one of the tasks running on its own would be fine, Its the shear load, our mmpmon data says the storage has been flat lining when it occurs.
Its a reasonably standard (small) HPC cluster, with a very mixed work load, hence while we can usually find "bad" jobs from the point of view of io on this occasion we can see a few large array jobs all accessing the same file, the cluster runs fine until we get to a certain point and one more will tip the balance. We've been attempting to limit the problem by adding limits to the number of jobs in an array that can run at once. But that feels like fire fighting.
Are you using GPFS API over any administrative commands? Any problems with the network (being that Ethernet or IB)?
We're not as using the GPFS API, never got it working, which is a shame, I've never managed to figure out the setup, although it is on my to do list.
Network wise, We've just removed a great deal of noise from arp requests by increasing the arp cache size on the nodes. Its a mixed 1GBit/10GBit network currently, we're currently looking at removing all the 1GBit nodes within the next few months and adding some new faster kit. The Storage is attached at 40GBit but it does not look to want to run much above 5Gbit I suspect due to Ethernet back off due to the mixed speeds.
While we do have some IB we don't currently run our storage over it.
Thanks in advance
Peter Childs
Sorry if I am un-announced here for the first time. But I would like to help if I can.
Jose Higino,
from NIWA
New Zealand
Cheers
On Sun, 22 Jul 2018 at 23:26, Peter Childs <p.childs at qmul.ac.uk<mailto:p.childs at qmul.ac.uk>> wrote:
Yes, we run mmbackup, using a snapshot.
The scan usally takes an hour, but for the last week has been taking many hours (i saw it take 12 last Tuesday)
It's speeded up again now back to its normal hour, but the high io jobs accessing the same file from many nodes also look to have come to an end for the time being.
I was trying to figure out howto control the bad io using mmchqos, to prioritise certain nodes over others but had not worked out if that was possible yet.
We've only previously seen this problem when we had some bad disks in our storage, which we replaced, I've checked and I can't see that issue currently.
Thanks for the help.
Peter Childs
Research Storage
ITS Research and Teaching Support
Queen Mary, University of London
---- Yaron Daniel wrote ----
Hi
Do u run mmbackup on snapshot , which is read only ?
Regards
________________________________
Yaron Daniel 94 Em Ha'Moshavot Rd
Storage Architect – IL Lab Services (Storage) Petach Tiqva, 49527
IBM Global Markets, Systems HW Sales Israel
Phone: +972-3-916-5672
Fax: +972-3-916-5672
Mobile: +972-52-8395593
e-mail: yard at il.ibm.com<mailto:yard at il.ibm.com>
IBM Israel<http://www.ibm.com/il/he/>
From: Peter Childs <p.childs at qmul.ac.uk<mailto:p.childs at qmul.ac.uk>>
To: "gpfsug-discuss at spectrumscale.org<mailto:gpfsug-discuss at spectrumscale.org>" <gpfsug-discuss at spectrumscale.org<mailto:gpfsug-discuss at spectrumscale.org>>
Date: 07/10/2018 05:51 PM
Subject: [gpfsug-discuss] Same file opened by many nodes / processes
Sent by: gpfsug-discuss-bounces at spectrumscale.org<mailto:gpfsug-discuss-bounces at spectrumscale.org>
________________________________
We have an situation where the same file is being read by around 5000
"jobs" this is an array job in uge with a tc set, so the file in
question is being opened by about 100 processes/jobs at the same time.
Its a ~200GB file so copying the file locally first is not an easy
answer, and these jobs are causing issues with mmbackup scanning the
file system, in that the scan is taking 3 hours instead of the normal
40-60 minutes.
This is read only access to the file, I don't know the specifics about
the job.
It looks like the metanode is moving around a fair amount (given what I
can see from mmfsadm saferdump file)
I'm wondering if we there is anything we can do to improve things or
that can be tuned within GPFS, I'm don't think we have an issue with
token management, but would increasing maxFileToCache on our token
manager node help say?
Is there anything else I should look at, to try and attempt to allow
GPFS to share this file better.
Thanks in advance
Peter Childs
_______________________________________________
gpfsug-discuss mailing list
gpfsug-discuss at spectrumscale.org<http://spectrumscale.org>
http://gpfsug.org/mailman/listinfo/gpfsug-discuss
--
Peter Childs
ITS Research Storage
Queen Mary, University of London
_______________________________________________
gpfsug-discuss mailing list
gpfsug-discuss at spectrumscale.org<http://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
--
Peter Childs
ITS Research Storage
Queen Mary, University of London
_______________________________________________
gpfsug-discuss mailing list
gpfsug-discuss at spectrumscale.org
http://gpfsug.org/mailman/listinfo/gpfsug-discuss
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gpfsug.org/pipermail/gpfsug-discuss_gpfsug.org/attachments/20180723/5b9e8da9/attachment.htm>
More information about the gpfsug-discuss
mailing list