Nancy:Production: Difference between revisions

From Grid5000
Jump to navigation Jump to search
(45 intermediate revisions by 9 users not shown)
Line 1: Line 1:
{{Portal|User}}
{{Portal|User}}
{{Portal|User}}


= Introduction =
= Introduction =
The Nancy Grid'5000 site also hosts nodes for production use. Those nodes are:
The Nancy Grid'5000 site also hosts clusters for production use (including clusters with GPUs). See [[Nancy:Hardware]] for details.
* '''talc''', a 144-node Intel Xeon cluster, which achieves a 7360 GFlops performance peak
* '''mbi''', a 32-node Intel Xeon cluster, with 7 GPU (mbi-[19-26]) (Tesla Nvidia)
* '''graphique''', a 6-node Intel Xeon cluster, with 2 GPU per node (GTX 980 Nvidia)


The usage rules differ from the rest of Grid'5000:
The usage rules differ from the rest of Grid'5000:
* Advance reservations (<code>oarsub -r</code>) are not allowed (to avoid fragmentation). Only submissions (and reservations that start immediately) are allowed.
* Advance reservations (<code>oarsub -r</code>) are not allowed (to avoid fragmentation). Only submissions (and reservations that start immediately) are allowed.
* All Grid'5000 users can use those nodes, but it is expected that users outside of LORIA / Inria Nancy -- Grand Est will use their own local production resources in priority, and mostly use those resources for tasks that require Grid'5000 features. Examples of local production clusters are Tompouce (Saclay), Igrida (Rennes), Plafrim (Bordeaux), etc.
* All Grid'5000 users can use those nodes (provided they meet the conditions stated in [[Grid5000:UsagePolicy]]), but it is expected that users outside of LORIA / Inria Nancy -- Grand Est will use their own local production resources in priority, and mostly use those resources for tasks that require Grid'5000 features. Examples of local production clusters are Tompouce (Saclay), Igrida (Rennes), Plafrim (Bordeaux), etc.


= Using the resources =
= Using the resources =
== Getting an account ==
== Getting an account ==
Please use the [[Special:G5KRequestAccountUMS|request form here]].
Users from Nancy (LORIA/Inria Nancy Grand-Est) that want to access Grid'5000 primarily for a production usage must use that '''[[Special:G5KRequestAccountUMS|request form]]''' to open an account, like regular Grid'5000 users.


* The following fields must be filled as indicated (but the '''other fields must be filled too'''):  
* The following fields must be filled as follows:
** manager: lnussbaum
** ''Group Granting Access'': either the group '''named after the research team''', or if it does not exist: '''<code>nancy-misc</code>'''.
** site: nancy
** ''Laboratory'': LORIA
** groups, roles: none
** ''Team'': SYNALP, MULTISPEECH, CARAMBA, CAPSID, ORPAILLEUR, etc.
** privileges: user
* Users are automatically subscribed to the Grid'5000 users mailing lists: [mailto:users@lists.grid5000.fr users@lists.grid5000.fr]. This list is the user-to-user or user-to-admin communication mean to address help/support requests for Grid'5000.
* You are automatically subscribed to the Grid'5000 users' mailing lists: users@lists.grid5000.fr
This list is the user-to-user or user-to-admin communication mean to address help/support requests for Grid'5000.


== Learning to use Grid'5000 ==
== Learning to use Grid'5000 ==
Refer to the [[Getting Started]] tutorial.
Refer to the [[Getting Started]] tutorial.
There are other tutorial listed on the [https://www.grid5000.fr/mediawiki/index.php/Category:Portal:User Users Home] page.
There are other tutorial listed on the [https://www.grid5000.fr/mediawiki/index.php/Category:Portal:User Users Home] page.
== Using deep learning software on Grid'5000 ==
A tutorial for using deep learning software on Grid'5000, written by Ismael Bada [[User:Ibada/Tuto_Deep_Learning|is also available]].


== Using production resources ==
== Using production resources ==
To access production resources, you need to submit jobs in the ''production'' queue:
To access production resources, you need to submit jobs in the ''production'' queue or using the ''production'' job type:
  oarsub -q production -I
  oarsub -q production -I
  oarsub -q production -p "cluster='talc'" -I
  oarsub -q production -p "cluster='grele'" -I
  oarsub -q production -l nodes=2,walltime=240 -I
  oarsub -q production -l nodes=2,walltime=24 -I
  oarsub -q production -l walltime=24 -t deploy 'sleep 100d'
  oarsub -q production -l walltime=24 -t deploy 'sleep 100d'
  ...
  ...
or
oarsub -t production -I
oarsub -t production -p "cluster='grele'" -I
oarsub -t production -l nodes=2,walltime=24 -I
oarsub -t production -l walltime=24 -t deploy 'sleep 100d'


== Dashboards and status pages ==
== Dashboards and status pages ==
Line 43: Line 47:
= Contact information and support =
= Contact information and support =
Contacts:
Contacts:
* The local system administrator is Clément Parisot ([mailto:clement.parisot@inria.fr clement.parisot@inria.fr])
* The Grid'5000 team can be contacted as described on the [[Support]] page.
* The Grid'5000 ''responsable de site'' for Nancy is Lucas Nussbaum ([mailto:lucas.nussbaum@loria.fr lucas.nussbaum@loria.fr])
* The Grid'5000 ''responsable de site'' for Nancy is Lucas Nussbaum ([mailto:lucas.nussbaum@loria.fr lucas.nussbaum@loria.fr])
* The rest of the Grid'5000 team can be contacted as described on the [[Support]] page.
* Ismael Bada (engineer funded by CPER LCHN) can also help local users, especially regarding requests related to deep learning on Grid'5000. ([mailto:ismael.bada@loria.fr ismael.bada@loria.fr])
 
To get support, you can:
To get support, you can:
* Use the [mailto:users@lists.grid5000.fr users@lists.grid5000.fr] mailing list: all Grid'5000 users (700+ people) are automatically subscribed
* Use the [mailto:users@lists.grid5000.fr users@lists.grid5000.fr] mailing list: all Grid'5000 users (700+ people) are automatically subscribed
* Use the [mailto:nancy-users@lists.grid5000.fr nancy-users@lists.grid5000.fr] mailing list: all Grid'5000 users from Nancy are automatically subscribed
* Use the [mailto:nancy-users@lists.grid5000.fr nancy-users@lists.grid5000.fr] mailing list: all Grid'5000 users from Nancy are automatically subscribed
* Contact Ismael Bada (see above)


The Grid'5000 team does not have the resources (manpower) to do user support, such as helping with writing scripts, creating system images, etc. If you need such help, please contact the SED service.
The Grid'5000 team does not have the resources (manpower) to do user support, such as helping with writing scripts, creating system images, etc. If you need such help, please contact either Ismael Bada (see above), or the SED service.


= FAQ =
= FAQ =
== Data storage ==
== Data storage ==
{{Warning|text=There is currently some problems with the storage on Grid'5000 production. The following informations will be updated frequently}}
The data needed for experiments of the production teams is stored on the talc-data and talc-data2 NFS servers. talc-data2 is a regular [[Group Storage]] server, and talc-data is a storage server dedicated to the multispeech research team, but compatible with the Group Storage mechanisms.
 
talc-data2 provides 213T of storage space. talc-data contains three disk bays (baie, baie2 and baie3) respectively hosting volumes talc, talc2 and talc2, respectively providing 58T + 58T + 71T = 187T of storage space.
 
  LV          VG        Attr      LSize  Pool Origin Data%  Meta%  Move Log Cpy%Sync Convert
  talc        baie      -wi-ao----  58.21t                                                   
  talc2        baie2    -wi-ao----  58.21t                                                   
  talc3        baie3    -wi-ao----  71.28t
 
/dev/mapper/baie2-talc2 on /export/group/talc2 type ext4 (rw,relatime,data=ordered)
/dev/mapper/baie3-talc3 on /export/group/talc3 type ext4 (rw,relatime,data=ordered)
/dev/mapper/baie-talc on /export/group/talc type ext4 (rw,relatime,stripe=32752,data=ordered)
 
Those repositories have the same permissions as any Group Storage volumes on their root folder,


Data we saved by successives reboot using rsync can be find on nancy frontend (fnancy) at /data/ftalc2_backup. It is mounted read-only.
drwxrws--T+ 4 root sto-multispeech 4096 Oct 13 22:13 talc
drwxrws--T+ 4 root sto-multispeech 4096 Mar 14  2019 talc2
drwxrws--T+ 4 root sto-multispeech 4096 Oct 10 10:50 talc3


For users that asked for data save, please check that what you asked was saved correctly.
but unlike other [[Group Storage]] volumes, they contain old data which is not affected by those permissions. Only the newly created directories under the root folder inherit those permissions.


If anyone forgot to signal very important data on /data1 or /home, let us know as soon as possible!
Please remember that those data are hosted on a NFS server that is not recommended for compute usage.


As we still don't know how much time it will take to finish backup and to bring nfs back, talc nodes have been integrated at Grid'5000 and can be used from fnancy using production queue.
For other shorter term storage support, see [[Storage]].


== I am physically located in the LORIA building, is there a shorter path to connect? ==
== I am physically located in the LORIA building, is there a shorter path to connect? ==
Line 69: Line 88:
If for some reason you don't want to go through Grid'5000 national access machines (access-south and access-north), you can also connect directly using
If for some reason you don't want to go through Grid'5000 national access machines (access-south and access-north), you can also connect directly using
{{Term|location=mylaptop|cmd=<code class="command">ssh</code> <code class="replace">jdoe</code><code>@</code><code class="host">access.nancy.grid5000.fr</code>}}
{{Term|location=mylaptop|cmd=<code class="command">ssh</code> <code class="replace">jdoe</code><code>@</code><code class="host">access.nancy.grid5000.fr</code>}}
== I have a large amount of jobs to execute, is there a better solution ? ==
Yes. You should have a look at [[CiGri]]. This middleware makes it easier to submit large number of jobs as ''best-effort' jobs on Grid'5000. The advantage is that you would benefit from the computing power of the whole of Grid'5000 for your jobs, not just of Nancy's production resources.


== How to access data in Inria/Loria ==
== How to access data in Inria/Loria ==
{{Warning|text=There is currently a bug that block this feature.
}}
bastionssh.loria.fr is an access machine hosted on Loria side.
bastionssh.loria.fr is an access machine hosted on Loria side.
That machine can be used to access all services in the Inria/Loria environment.
That machine can be used to access all services in the Inria/Loria environment.


You need to use SSH ProxyCommand for that purpose.
You need to use the SSH ProxyCommand for that purpose.


Adapt following lines for your <code class="file">~/.ssh/config</code>
Ajust the following lines for your <code class="file">~/.ssh/config</code>
<pre>
<pre>
  Host accessloria
Host accessloria
         HostName bastionssh.loria.fr
         Hostname bastionssh.loria.fr
         User jdoe
         User jdoe # to be replaced by your LORIA login
        Port 22
 
        IdentityFile /home/%u/.ssh/id_rsa
Host *.loria
        ForwardAgent no
         User jdoe # to be replaced by your LORIA login
  Host *.loria
         User jdoe
         ProxyCommand ssh accessloria -W $(basename %h .loria):%p
         ProxyCommand ssh accessloria -W $(basename %h .loria):%p
        IdentityFile /home/%u/.ssh/id_rsa
</pre>         
</pre>         


Line 101: Line 112:
eg:
eg:


{{Term|location=fnancy|cmd=<code class="command">sshfs</code> <code>-o idmap=user</code> <code class="replace">jdoe</code><code>@</code><code class="host">tregastel.loria</code>:<code class="file">/users/myteam/jdoe ~/remote_dir</code>}}
{{Term|location=fnancy|cmd=<code class="command">sshfs</code> <code>-o idmap=user</code> <code class="replace">jdoe</code><code>@</code><code class="host">tregastel.loria</code>:<code class="file">/users/myteam/jdoe ~/local_dir</code>}}
    
    
To unmount the remote filesystem:  
To unmount the remote filesystem:  
{{Term|location=fnancy|cmd=<code class="command">fusermount</code> <code>-u</code> <code class="file">~/remote_dir</code>}}
{{Term|location=fnancy|cmd=<code class="command">fusermount</code> <code>-u</code> <code class="file">~/local_dir</code>}}
    
    
{{Note|text=The password for your Inria/Loria account will be asked twice if you don't specify a ssh key.  
{{Note|text=Given that bastionssh.loria.fr only accepts logins using SSH key, you cannot simply connect with your LORIA password.}}
}}


== I submitted a job, there are free resources, but my job doesn't start as expected! ==
== I submitted a job, there are free resources, but my job doesn't start as expected! ==
Most likely, this is because of our configuration of resources restriction per walltime.
Most likely, this is because of our configuration of resources restriction per walltime.
In order to make sure that someone requesting only a few nodes, for a small amount of time will be able to get soon enough, the nodes are split into categories:
In order to make sure that someone requesting only a few nodes, for a small amount of time will be able to get soon enough, the nodes are split into categories. This depends on each cluster and is visible in the Gantt chart. An example split is:
* 20% of the nodes only accept jobs with walltime lower than 1h
* 20% of the nodes only accept jobs with walltime lower than 1h
* 20% -- 2h
* 20% -- 2h
* 20% -- 24h
* 20% -- 24h
* 20% -- 48h
* 20% -- 48h
* 20% accept all jobs (no limit on duration)
* 20% -- one week
Note that ''best-effort'' jobs are excluded from those limitations.
Note that ''best-effort'' jobs are excluded from those limitations.


Line 123: Line 133:
== I have an important demo, can I reserve all resources in advance? ==
== I have an important demo, can I reserve all resources in advance? ==
There's a special ''challenge'' queue that can be used to combine resources from the classic Grid'5000 clusters and the production clusters for special events. If you would like to use it, please get in touch with the clusters managers.
There's a special ''challenge'' queue that can be used to combine resources from the classic Grid'5000 clusters and the production clusters for special events. If you would like to use it, please get in touch with the clusters managers.
==Can I use besteffort jobs in production ?==
Yes, you can submit a besteffort job on the production resources by using OAR <code>-t besteffort</code> option. Here is an exemple:
{{Term|location=fnancy|cmd=<code class="command">oarsub</code> <code class="replace">-t besteffort</code> <code>-q production</code><code class="file">./my_script.sh</code>}}
If you didn't specify the <code>-q production</code> option, your job could run on both production and non-production resources.
== Is it possible to run Matlab? ==
Yes, through SSH tunneling to access UL license server (access to bastionssh.loria.fr required).
More information is available in [https://members.loria.fr/FSur/articles/MatlabGrid5000.pdf this document].


== Energy costs ==
== Energy costs ==
Line 142: Line 161:
** Heure Pleine Eté 4,125c€/KWh
** Heure Pleine Eté 4,125c€/KWh
** Heure Creuse Eté 2,580c€/KWh
** Heure Creuse Eté 2,580c€/KWh
== How to cite / Comment citer ==
If you use the Grid'5000 production clusters for your research and publish your work, please add this sentence in the acknowledgements section of your paper:
<blockquote>
Experiments presented in this paper were carried out using the Grid'5000
testbed, supported by a scientific interest group hosted by
Inria and including CNRS, RENATER and several Universities as well as
other organizations (see https://www.grid5000.fr).
</blockquote>

Revision as of 08:51, 12 October 2020


Introduction

The Nancy Grid'5000 site also hosts clusters for production use (including clusters with GPUs). See Nancy:Hardware for details.

The usage rules differ from the rest of Grid'5000:

  • Advance reservations (oarsub -r) are not allowed (to avoid fragmentation). Only submissions (and reservations that start immediately) are allowed.
  • All Grid'5000 users can use those nodes (provided they meet the conditions stated in Grid5000:UsagePolicy), but it is expected that users outside of LORIA / Inria Nancy -- Grand Est will use their own local production resources in priority, and mostly use those resources for tasks that require Grid'5000 features. Examples of local production clusters are Tompouce (Saclay), Igrida (Rennes), Plafrim (Bordeaux), etc.

Using the resources

Getting an account

Users from Nancy (LORIA/Inria Nancy Grand-Est) that want to access Grid'5000 primarily for a production usage must use that request form to open an account, like regular Grid'5000 users.

  • The following fields must be filled as follows:
    • Group Granting Access: either the group named after the research team, or if it does not exist: nancy-misc.
    • Laboratory: LORIA
    • Team: SYNALP, MULTISPEECH, CARAMBA, CAPSID, ORPAILLEUR, etc.
  • Users are automatically subscribed to the Grid'5000 users mailing lists: users@lists.grid5000.fr. This list is the user-to-user or user-to-admin communication mean to address help/support requests for Grid'5000.

Learning to use Grid'5000

Refer to the Getting Started tutorial. There are other tutorial listed on the Users Home page.

Using deep learning software on Grid'5000

A tutorial for using deep learning software on Grid'5000, written by Ismael Bada is also available.

Using production resources

To access production resources, you need to submit jobs in the production queue or using the production job type:

oarsub -q production -I
oarsub -q production -p "cluster='grele'" -I
oarsub -q production -l nodes=2,walltime=24 -I
oarsub -q production -l walltime=24 -t deploy 'sleep 100d'
...

or

oarsub -t production -I
oarsub -t production -p "cluster='grele'" -I
oarsub -t production -l nodes=2,walltime=24 -I
oarsub -t production -l walltime=24 -t deploy 'sleep 100d'

Dashboards and status pages

Contact information and support

Contacts:

  • The Grid'5000 team can be contacted as described on the Support page.
  • The Grid'5000 responsable de site for Nancy is Lucas Nussbaum (lucas.nussbaum@loria.fr)
  • Ismael Bada (engineer funded by CPER LCHN) can also help local users, especially regarding requests related to deep learning on Grid'5000. (ismael.bada@loria.fr)

To get support, you can:

The Grid'5000 team does not have the resources (manpower) to do user support, such as helping with writing scripts, creating system images, etc. If you need such help, please contact either Ismael Bada (see above), or the SED service.

FAQ

Data storage

The data needed for experiments of the production teams is stored on the talc-data and talc-data2 NFS servers. talc-data2 is a regular Group Storage server, and talc-data is a storage server dedicated to the multispeech research team, but compatible with the Group Storage mechanisms.

talc-data2 provides 213T of storage space. talc-data contains three disk bays (baie, baie2 and baie3) respectively hosting volumes talc, talc2 and talc2, respectively providing 58T + 58T + 71T = 187T of storage space.

 LV           VG        Attr       LSize   Pool Origin Data%  Meta%  Move Log Cpy%Sync Convert
 talc         baie      -wi-ao----  58.21t                                                    
 talc2        baie2     -wi-ao----  58.21t                                                    
 talc3        baie3     -wi-ao----  71.28t
/dev/mapper/baie2-talc2 on /export/group/talc2 type ext4 (rw,relatime,data=ordered)
/dev/mapper/baie3-talc3 on /export/group/talc3 type ext4 (rw,relatime,data=ordered)
/dev/mapper/baie-talc on /export/group/talc type ext4 (rw,relatime,stripe=32752,data=ordered)

Those repositories have the same permissions as any Group Storage volumes on their root folder,

drwxrws--T+ 4 root sto-multispeech 4096 Oct 13 22:13 talc
drwxrws--T+ 4 root sto-multispeech 4096 Mar 14  2019 talc2
drwxrws--T+ 4 root sto-multispeech 4096 Oct 10 10:50 talc3

but unlike other Group Storage volumes, they contain old data which is not affected by those permissions. Only the newly created directories under the root folder inherit those permissions.

Please remember that those data are hosted on a NFS server that is not recommended for compute usage.

For other shorter term storage support, see Storage.

I am physically located in the LORIA building, is there a shorter path to connect?

If for some reason you don't want to go through Grid'5000 national access machines (access-south and access-north), you can also connect directly using

Terminal.png mylaptop:
ssh jdoe@access.nancy.grid5000.fr

How to access data in Inria/Loria

bastionssh.loria.fr is an access machine hosted on Loria side. That machine can be used to access all services in the Inria/Loria environment.

You need to use the SSH ProxyCommand for that purpose.

Ajust the following lines for your ~/.ssh/config

Host accessloria
        Hostname bastionssh.loria.fr
        User jdoe # to be replaced by your LORIA login

Host *.loria
        User jdoe # to be replaced by your LORIA login
        ProxyCommand ssh accessloria -W $(basename %h .loria):%p

With that setup, you can now use :

  • Rsync to synchronize your data on Inria/Loria environment and data on your local home on Grid'5000 frontend
  • Sshfs to mount directly your data directory on Inria/Loria environment under your local home. <=> mount your /user/my_team/my_username (origin = bastionssh.loria.fr) on fnancy (destination = a folder on fnancy).

eg:

Terminal.png fnancy:
sshfs -o idmap=user jdoe@tregastel.loria:/users/myteam/jdoe ~/local_dir

To unmount the remote filesystem:

Terminal.png fnancy:
fusermount -u ~/local_dir
Note.png Note

Given that bastionssh.loria.fr only accepts logins using SSH key, you cannot simply connect with your LORIA password.

I submitted a job, there are free resources, but my job doesn't start as expected!

Most likely, this is because of our configuration of resources restriction per walltime. In order to make sure that someone requesting only a few nodes, for a small amount of time will be able to get soon enough, the nodes are split into categories. This depends on each cluster and is visible in the Gantt chart. An example split is:

  • 20% of the nodes only accept jobs with walltime lower than 1h
  • 20% -- 2h
  • 20% -- 24h
  • 20% -- 48h
  • 20% -- one week

Note that best-effort jobs are excluded from those limitations.

Another enabled OAR feature that could impact the scheduling of your jobs is the OAR karma: this feature assigns a dynamic priority to submissions based on the history of submissions by a specific user. With that feature, the jobs from users that rarely submit jobs will be generally scheduled earlier than jobs from heavy users.

I have an important demo, can I reserve all resources in advance?

There's a special challenge queue that can be used to combine resources from the classic Grid'5000 clusters and the production clusters for special events. If you would like to use it, please get in touch with the clusters managers.

Can I use besteffort jobs in production ?

Yes, you can submit a besteffort job on the production resources by using OAR -t besteffort option. Here is an exemple:

Terminal.png fnancy:
oarsub -t besteffort -q production./my_script.sh

If you didn't specify the -q production option, your job could run on both production and non-production resources.

Is it possible to run Matlab?

Yes, through SSH tunneling to access UL license server (access to bastionssh.loria.fr required). More information is available in this document.

Energy costs

Grid'5000 nodes are automatically shut down when they are not reserved so, when possible, it is a good idea to reserve nodes during cheaper time slots.

Electricity costs are currently:

  • Périodes:
    • Heures pointe: Décembre, Janvier, Février; 09H00-11H00/18H00-20H00
    • Heures Pleines Hiver: 06H00-22H00 (hors heures de pointe précisées à l’article 18).
    • Heures Creuses Hiver: 22H00-06H00
    • Heures Pleines Eté: 06H00-22H00
    • Heures Creuses Eté: 22H00-06H00
    • Le dimanche ne comprend que des heures creuses en hiver et été.
  • Cout du KWh
    • Heure pointe 10,893c€/KWh
    • Heure Pleine Hiver 6,535c€/KWh
    • Heure Creuse Hiver 4,474c€/KWh
    • Heure Pleine Eté 4,125c€/KWh
    • Heure Creuse Eté 2,580c€/KWh

How to cite / Comment citer

If you use the Grid'5000 production clusters for your research and publish your work, please add this sentence in the acknowledgements section of your paper:

Experiments presented in this paper were carried out using the Grid'5000 testbed, supported by a scientific interest group hosted by Inria and including CNRS, RENATER and several Universities as well as other organizations (see https://www.grid5000.fr).