[Beowulf] Most common cluster management software, job schedulers, etc?

[Prentice Bisbal]

On 03/08/2016 11:16 AM, Remy Dernat wrote:
> Hi,
>
> Le 08/03/2016 09:25, Carsten Aulbert a écrit :
>> Hi
>>
>> On 03/08/2016 05:43 AM, Jeff Friedman wrote:
>>> Hello all. I am just entering the HPC Sales Engineering role, and would
>>> like to focus my learning on the most relevant stuff. I have searched
>>> near and far for a current survey of some sort listing the top used
>>> “stacks”, but cannot seem to find one that is free. I was breaking
>>> things down similar to this:
>> "relevant" stuff is pretty relative to what you want to achieve ;)
>>> _Provisioning software_: Cobbler, Warewulf, xCAT, Openstack, 
>>> Platform HPC, ?
> Well, OpenStack is designed for cloud, not for HPC, but perhaps some 
> people are using OpenStack in that purpose...
>
> You could add RocksCluster, sidus ( 
> http://www.cbp.ens-lyon.fr/doku.php?id=en:developpement:productions:sidus 
> ), kadeploy ( http://kadeploy3.gforge.inria.fr/ ), perceus ( 
> http://moo.nac.uci.edu/~hjm/Perceus-Report.html )...
>>>
>> In case of Debian: FAI
>
> You could also use FAI to serve non-debian-like systems. I use it to 
> deploy ubuntu but you can also deploy redhat-like system, even if it 
> is quite harder. Only the first boot system (through DHCP/PXE and 
> then, NFS) is debian (nfsroot), then it can install what you need.
>>
>>> _Configuration management_: Warewulf, Puppet, Chef, Ansible, ?
>>>
> + SaltStack ?
>
> Generally, people are not using that kind of stuff in HPC, but yes, it 
> could happen.

Says you! ;)

I used to just do my cluster configuration using a postinstall script in 
Kickstart (as you mention below), but once I started using Puppet for my 
non-cluster systems, it made little sense to use two different 
configuration management methodologies within the enterprise, so I 
switched to just calling 'puppet agent' from the postinstall script.  
The only difference is that to reduce overhead, I don't keep the puppet 
agent daemons running on the compute nodes, I used gsh to run 'puppet 
agent' on-demand. Nowadays, I'd use pdsh instead of gsh.
>
> You have some images to install the whole cluster, and eventually a 
> post configuration step (kickstart ?), if you have a diskfull 
> configuration.
>
> Then, you can use cluster-command tool like cluster-ssh, pssh, pdsh...
>> old school: cfengine, ...
>>
>>> _Resource and job schedulers_: I think these are basically the same
>>> thing? Torque, Lava, Maui, Moab, SLURM, Grid Engine, Son of Grid 
>>> Engine,
>>> Univa, Platform LSF, etc… others?
> + OAR https://oar.imag.fr/
>
> Some Job Schedulers are also Resource Manager, but it is not always 
> true : 
> https://wiki.hpcc.msu.edu/display/hpccdocs/Resource+Managment+and+Job+Scheduler
>
>> for high throughput computing: HTCondor
>>
>>> _Performance monitoring_: Ganglia, Nagios, ?
>> Icinga, ...
> Shinken, zabbix, etc... There are also some new tools with others 
> storage and display technology (influxDB , graphite, grafana...)...
>
> But for HPC, ganglia is good enough...
>
> Best,
> Remy.
>
> PS : good to know this "small hpc" google site.
>>
>>> Does anyone have any observations as to which of the above are the most
>>> common?  Or is that too broad?  I  believe most the clusters I will be
>>> involved with will be in the 128 - 2000 core range, all on commodity
>>> hardware.
>> I guess everyone will have their preferences, if you wanted to get to
>> some hard, recent numbers, one way would be to crate an online
>> survey/form and ask many people to participate :)
>>
>> Cheers
>>
>> Carsten
>> _______________________________________________
>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
>> To change your subscription (digest mode or unsubscribe) visit 
>> http://www.beowulf.org/mailman/listinfo/beowulf
>>
>