I have configured the SchedulerType=sched/backfill
with 30 node in the partition
gpu-1-4                     1 batch*    mixed         32    2:8:2   258433    0 
 2051220 gtx680,mat none
gpu-1-[5-7]                 3 batch*    idle          32    2:8:2  258433+    0 
 2051220 gtx680,mat none
gpu-1-[8-10]                3 batch*    idle          32    2:8:2  258433    0  
2051220 gtx680,mat none
gpu-1-11                    1 batch*    idle          32    2:8:2  258433    0  
2051220 gtx680,mat none
gpu-1-[12-13]               2 batch*    idle          32    2:8:2   258423    0 
 2051220 32CPUs,mat none
gpu-1-14                    1 batch*    idle          32    2:8:2   258433    0 
 2051220 gtx680,mat none
gpu-1-16,gpu-2-[10-17]      9 batch*    idle          32    2:8:2  258423+    0 
 2051220 gtxtitan   none
gpu-1-17,gpu-2-[4-9],gpu-   8 batch*    idle          32    2:8:2  258423+    0 
 2051220 gtx680     none


slurm uses one sinlge node for the 4 jobs while all other nodes are
idle.

# squeue
             JOBID PARTITION     NAME PRIORITY     USER  STATE
TIME  NODES  CPUS NODELIST(REASON)
            172048     batch Exp.1-81 0.000000  akahles RUNNIN       0:16      
1     2 gpu-1-4
            172047     batch Exp.1-81 0.000000  akahles RUNNIN       0:22      
1     2 gpu-1-4
            172046     batch rDi_TCGA 0.000000  akahles RUNNIN       6:51      
1     2 gpu-1-4
            172045     batch rDi_TCGA 0.000000  akahles RUNNIN       6:56      
1     2 gpu-1-4

I have set the node weight factor to the same value on all nodes but the
weight never changes, no matter if the node is busy or not. Is there
something like a priority factor in slurm which gets calculated by the
node usage and the least used node would be used?

How do I configure slurm to allocate a new node per job if available?

Thanks
Eva

Reply via email to