This guide explains how purchased resources are allocated on our clusters.

There are 3 different types of account on our clusters and the full description can be found here

Free

Free (or freemium) accounts cost nothing but allow limited access to the resources. The main limits are that the maximum run time of a job is 24 hours and that only one node can be user per job. Jobs submitted by free users will also be scheduled with a lower priority than those of premium or share users.

Premium

Premium accounts are charged for their real usage of the system along with a monthly administration fee. The maxmium wall time is 3 days (extendable upon request) and 8 nodes per job (1 on Castor). Jobs from Premium users are scheduled with an elevated priority compared to those from free.

Please note that Premium is not available on Bellatrix for capacity reasons.

To move from a Free to a Premium account please ask the head of the laboratory to contact SCITAS at 1234@epfl.ch

Share

Shares can be purchased either for the lifetime of a machine or on a semestrial basis. Here the account is assigned a weight equivalent to the number of CPUs they have purchased and the scheduler aims to ensure that this amount is allocated. In the case of there being idle nodes it is possible to use more than the number of CPUs purchased. After the initial purchase there are no further costs.

 

Fair-share

 

The fair-share algorithm in SLURM is described at http://slurm.schedmd.com/fair_tree.html and we use a half-life of one week. 

To see the share for your group you can use the "Sshare" command"

$ Sshare 

             Account       User Raw Shares Norm Shares   Raw Usage  Norm Usage Effectv Usage  FairShare   Level FS  
-------------------- ---------- ---------- ----------- ----------- ----------- ------------- ---------- ---------- 
scitas-ge                                1    0.007752        1376    0.000003      0.000005           1468.763590
 scitas-ge               aubort          1    0.043478           0    0.000000      0.000000   0.290000        inf 
 scitas-ge             clemenco          1    0.043478           0    0.000000      0.000000   0.290000        inf
 scitas-ge                cubuk          1    0.043478           0    0.000000      0.000000   0.290000        inf
 scitas-ge                culpo          1    0.043478           0    0.000000      0.000000   0.290000        inf
 scitas-ge             degiorgi          1    0.043478           0    0.000000      0.000000   0.290000        inf
 scitas-ge               eroche          1    0.043478         344    0.000001      0.250000   0.253333   0.173913
 scitas-ge              nvarini          1    0.043478           0    0.000000      0.000000   0.290000        inf
 scitas-ge                qubit          1    0.043478         351    0.000001      0.255072   0.250000   0.170455
 scitas-ge             rezzonic          1    0.043478         681    0.000001      0.494928   0.246667   0.087848
 scitas-ge              richart          1    0.043478           0    0.000000      0.000000   0.290000        inf
 scitas-ge              rmsilva          1    0.043478           0    0.000000      0.000000   0.290000        inf
 scitas-ge                  sue          1    0.043478           0    0.000000      0.000000   0.290000        inf
 scitas-ge                 topf          1    0.043478           0    0.000000      0.000000   0.290000        inf
 

The "Norm Shares" column on the first line is the proportion of the cluster which is allocated to the account and the shares are in terms of cores. Within a group all users have equal weight and so 1 share each.

 

The value used to decide the priority of a job is the "Level FS" and this is calculated based on the difference between the "Norm Shares" and "Effectv Usage" values. The higher the Level FS, the higher the priority.