[slurm-dev] Re: How to distinguish slurmctld.pid of server and backup-server on a shared disk(slurm 14.11)
Hi, Chrysovalantis. Thanks a lot for the prompt reply. Actually we have many nodes and put all the logs/pids/tmpfiles/ in one folder named slurm-xxx on a shared filesystem. My boss think it's a good idea because we won't have to update every node on every update attempt. Well, I think it's not bad either. Hah. Simbolic links is a feasible solution. But I will have to remove the local simbolic link once some nodes is to be removed for any other use. Now I put the pid files into /var/run and it will be deleted if the process is killed. And the logs of slurmctlds of server and backup server will be put in one file. Anyway, thanks a lot for your kind help. May you have a good day. 2015-06-10 18:57 GMT+08:00 Chrysovalantis Paschoulas < c.paschou...@fz-juelich.de>: > Hi! > > In our site we store the logs and the pid files of slurmctld and slurmdbbd > daemons on local disk, it is better so to distinguish who is doing what and > since if a node is down the daemons are not running then there is no need > to put them on a shared filesystem. For slurmctld we have the state dir and > jobcomp on a shared active/active filesystem. For slurmdbd we have an > active/passive filesystem between the master nodes where we store the > mysql/mariadb database. > > I understand that it is your policy to use shared filesystems for those > files (and in general is not a bad idea) but in this case for slurm control > daemons you should use local disks (personal opinion). > > Anyway, if you insist to use shared fs then I can only think of one > solution: use symbolic links. E.g.: > @master1: path-of-pidfile-in-slurm.conf -> /sharedfs/slurm/run/master1.pid > @master2: path-of-pidfile-in-slurm.conf -> /sharedfs/slurm/run/master2.pid > > Cheers, > Chrysovalantis Paschoulas > > > > On 06/10/2015 04:47 AM, Qianqian Sha wrote: > > Hi, > > We store all slurm logs/pids/tmpfiles on a shared disk. > logs/pids/tmpfiles of different slurmds can be distinguished by > nodename(%n) or hostname(%h). But it seems that logs/pids path of slurmctld > or backup-slurmctld does not support %n or %h. > > Any assistance or advice is appreciated. > > Thanks. > > > > > > > > > Forschungszentrum Juelich GmbH > 52425 Juelich > Sitz der Gesellschaft: Juelich > Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498 > Vorsitzender des Aufsichtsrats: MinDir Dr. Karl Eugen Huthmacher > Geschaeftsfuehrung: Prof. Dr.-Ing. Wolfgang Marquardt (Vorsitzender), > Karsten Beneke (stellv. Vorsitzender), Prof. Dr.-Ing. Harald Bolt, > Prof. Dr. Sebastian M. Schmidt > > > > > >
[slurm-dev] What is the reload function in /etc/init.d/slurm supposed to do?
I had a bit of an adventure with that - I accidentally brought all the slurm nodes to a "not responding" state (no job interruption). I'm mid upgrade from 2.6.1 to 14.11.7 - so transient state, no call to fix anything, just a question. After my adventure, I realized I'm not sure what is really supposed to happen on 'reload'. It looks like the controller asks all the nodes to reload their configs? Didn't find anything in the docs. Here are the details...more as an fyi and entertainment: I upgraded the dbd, controller, and one compute node - no problems - using a pre-existing init.d/slurm script. I'm improving the Debian package for internal distribution and see that the init.d scripts in the distribution do not support Debian. ( 1. location of init functions is /lib/lsb/init-functions and 2. and mandatory support of force_reload() - I'll be happy to add to the dev stream) I'm also newish to slurm, so I'm trying to gain a deep understanding of the init.d script to learn about slurm. Investigating how to implement force_reload(), I see in /etc/init.d/slum that 'reload' equates to: killproc $prog -HUP I think, "fine, I can reload the configs instead of restarting the service every time!" So I do a: kill -1 #-- that's kill-dash-one At first, the controller log shows: [2015-06-11T07:44:12.224] Reconfigure signal (SIGHUP) received [2015-06-11T07:44:12.235] restoring original state of nodes [2015-06-11T07:44:12.235] restoring original partition state [2015-06-11T07:44:12.246] cons_res: select_p_node_init [2015-06-11T07:44:12.246] cons_res: preparing for 5 partitions [2015-06-11T07:44:12.447] read_slurm_conf: backup_controller not specified. [2015-06-11T07:44:12.447] cons_res: select_p_reconfigure [2015-06-11T07:44:12.447] cons_res: select_p_node_init [2015-06-11T07:44:12.447] cons_res: preparing for 5 partitions Then the controller starts spitting out errors. [2015-06-11T07:45:15.228] agent/is_node_resp: node: rpc:1001 : Incompatible versions of client and server code One per node. Then all the nodes stop responding. error: Nodes not responding I restarted every node's slurmd and everything was back to normal. I guess I expected only the controller to reload its config file, but all the better that it asks all the nodes to do the same. I'll of course not do this again until the upgrade is complete. Thanks Sean -- Sean Blanton, Ph.D.
[slurm-dev] concurrent job limit
Is there a way for users to self limit the number of jobs that they concurrently run? Eric Martin Center for Genome Sciences & Systems Biology Washington University School of Medicine Forest Park Avenue St. Louis, MO 63108 The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail.
[slurm-dev] Re: What is the reload function in /etc/init.d/slurm supposed to do?
Minor correction, that's support for Ubuntu, and the missing directive is 'force-reload', with a dash, not an underscore. On Thu, Jun 11, 2015 at 8:36 AM, Sean Blanton wrote: > I had a bit of an adventure with that - I accidentally brought all the > slurm nodes to a "not responding" state (no job interruption). I'm mid > upgrade from 2.6.1 to 14.11.7 - so transient state, no call to fix > anything, just a question. > > After my adventure, I realized I'm not sure what is really supposed to > happen on 'reload'. It looks like the controller asks all the nodes to > reload their configs? Didn't find anything in the docs. > > Here are the details...more as an fyi and entertainment: > > I upgraded the dbd, controller, and one compute node - no problems - using > a pre-existing init.d/slurm script. I'm improving the Debian package for > internal distribution and see that the init.d scripts in the distribution > do not support Debian. ( 1. location of init functions is > /lib/lsb/init-functions and 2. and mandatory support of force_reload() - > I'll be happy to add to the dev stream) I'm also newish to slurm, so I'm > trying to gain a deep understanding of the init.d script to learn about > slurm. > > Investigating how to implement force_reload(), I see in /etc/init.d/slum > that 'reload' equates to: > > killproc $prog -HUP > > I think, "fine, I can reload the configs instead of restarting the service > every time!" So I do a: > > kill -1 #-- that's kill-dash-one > > At first, the controller log shows: > > [2015-06-11T07:44:12.224] Reconfigure signal (SIGHUP) received > [2015-06-11T07:44:12.235] restoring original state of nodes > [2015-06-11T07:44:12.235] restoring original partition state > [2015-06-11T07:44:12.246] cons_res: select_p_node_init > [2015-06-11T07:44:12.246] cons_res: preparing for 5 partitions > [2015-06-11T07:44:12.447] read_slurm_conf: backup_controller not specified. > [2015-06-11T07:44:12.447] cons_res: select_p_reconfigure > [2015-06-11T07:44:12.447] cons_res: select_p_node_init > [2015-06-11T07:44:12.447] cons_res: preparing for 5 partitions > > Then the controller starts spitting out errors. > > [2015-06-11T07:45:15.228] agent/is_node_resp: node: rpc:1001 : > Incompatible versions of client and server code > > One per node. Then all the nodes stop responding. > > error: Nodes not responding > > I restarted every node's slurmd and everything was back to normal. > > I guess I expected only the controller to reload its config file, but all > the better that it asks all the nodes to do the same. > > I'll of course not do this again until the upgrade is complete. > > > > Thanks > Sean > > -- > Sean Blanton, Ph.D. > > -- Sean Blanton, Ph.D. Quantitative Technologist Radix Trading, LLC Desk: 773.985.0456 Cell: 773.960.3495
[slurm-dev] Re: concurrent job limit
Eric, We find that the "honor system" frequently needs to be policed. We use SLURM's QOS and accounting systems to limit the number of concurrent cores, maximum running jobs, and maximum pending jobs. You may want to investigate this. With a QOS, you can apply limits per users (specifically cores), and with accounting you can apply limits per user or as a whole for a group (account). John DeSantis 2015-06-11 10:12 GMT-04:00 Martin, Eric : > Is there a way for users to self limit the number of jobs that they > concurrently run? > > Eric Martin > Center for Genome Sciences & Systems Biology > Washington University School of Medicine > Forest Park Avenue > St. Louis, MO 63108 > > > > > > The materials in this message are private and may contain Protected > Healthcare Information or other information of a sensitive nature. If you > are not the intended recipient, be advised that any unauthorized use, > disclosure, copying or the taking of any action in reliance on the contents > of this information is strictly prohibited. If you have received this email > in error, please immediately notify the sender via telephone or return mail.
[slurm-dev] Re: Question about running interactive job on all cores on a list of heterogeneous nodes.
Hi Jim, I think it may depend on how the "SelectType=“ is setup in your config. If it’s set as cons_res, then you should be able to do something like below: Sum up the number of total cores from the list of nodes before hand, maybe this isn’t possible though. Then do: srun -n -w node[1-50] Or srun -n -w /path/to/node_list_file Hope that helps. Best, Chris On 6/10/15, 2:26 PM, "Jim Robanske" wrote: >Hello, > >I've been trying to figure out how to run an interactive job on all cores >on a list of heterogeneous nodes (meaning that the nodes in the list may >have different numbers of cores). > >Anyone out there know how to accomplish this? > >Thanks in advance... > >-- >I may be inconsistent, but at least I'm consistently inconsistent.
[slurm-dev] Re: Question about running interactive job on all cores on a list of heterogeneous nodes.
Also try the --exclusive option on the job submit command line. Quoting Christopher B Coffey : Hi Jim, I think it may depend on how the "SelectType=“ is setup in your config. If it’s set as cons_res, then you should be able to do something like below: Sum up the number of total cores from the list of nodes before hand, maybe this isn’t possible though. Then do: srun -n -w node[1-50] Or srun -n -w /path/to/node_list_file Hope that helps. Best, Chris On 6/10/15, 2:26 PM, "Jim Robanske" wrote: Hello, I've been trying to figure out how to run an interactive job on all cores on a list of heterogeneous nodes (meaning that the nodes in the list may have different numbers of cores). Anyone out there know how to accomplish this? Thanks in advance... -- I may be inconsistent, but at least I'm consistently inconsistent. -- Morris "Moe" Jette CTO, SchedMD LLC Commercial Slurm Development and Support
[slurm-dev] Re: concurrent job limit
Job arrays can kind of be used for that: From http://slurm.schedmd.com/job_array.html: A maximum number of simultaneously running tasks from the job array may be specified using a "%" separator. For example "--array=0-15%4" will limit the number of simultaneously running tasks from this job array to 4. Ryan On 06/11/2015 08:12 AM, Martin, Eric wrote: Is there a way for users to self limit the number of jobs that they concurrently run? Eric Martin Center for Genome Sciences & Systems Biology Washington University School of Medicine Forest Park Avenue St. Louis, MO 63108 The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail. -- Ryan Cox Operations Director Fulton Supercomputing Lab Brigham Young University
[slurm-dev] slurm on systemd
Hi all, I was wondering if slurm run on systemd operating systems like fedora 22. I installed the tarball and right now I can't figure out how to start the slurm daemons. Does anyone know how to start the daemons in Fedora/if slurm is compatible with systemd? Thanks, Trevor
[slurm-dev] Re: slurm on systemd
On 12/06/15 09:43, Trevor Gale wrote: > I was wondering if slurm run on systemd operating systems like fedora > 22. I installed the tarball and right now I can't figure out how to > start the slurm daemons. Does anyone know how to start the daemons in > Fedora/if slurm is compatible with systemd? I did run it on CentOS 7 as a test a while ago, we never start slurm on boot and just logging in as usual and running our RHEL6 init script to start it worked fine. Of course that won't give you proper systemd integration, but it did seem to work in that way for simple testing at least. All the best, Chris -- Christopher SamuelSenior Systems Administrator VLSCI - Victorian Life Sciences Computation Initiative Email: sam...@unimelb.edu.au Phone: +61 (0)3 903 55545 http://www.vlsci.org.au/ http://twitter.com/vlsci
[slurm-dev] Re: Slurm and docker/containers
On 05/06/15 19:31, Nathan Harper wrote: > Has anyone looked at using LXC rather than Docker specifically? From > what I understand, it's possible to run unprivileged LXC containers, so > no need to be root. Speaking to some of the uni's OpenStack folks at lunch today about containerisation they recommended looking at LXD from Canonical which is an evolution of LXC. http://www.ubuntu.com/cloud/tools/lxd Of course getting it to run on anything other than Ubuntu might be a challenge which could limit its usefulness. All the best, Chris -- Christopher SamuelSenior Systems Administrator VLSCI - Victorian Life Sciences Computation Initiative Email: sam...@unimelb.edu.au Phone: +61 (0)3 903 55545 http://www.vlsci.org.au/ http://twitter.com/vlsci