Case Study: Node Board Scheduling with PBSPro 9.x and SGI Altix

A.27 Case Study: Node Board Scheduling with PBSPro 9.x and SGI Altix

Overview

Sites that have PBSPro 9.x and SGI Altix systems may find Moab node board scheduling capabilities useful. Moab can take a request of CPUs and memory and determine how many node boards are required for the job. For example, a user may request one processor and 14 GB of memory (qsub -l ncpus=1,mem=14gb script) on a node that has four processors and 8 GB of memory per node board. Moab would determine that the job would need two node boards to fulfill the request.

Configuration

RMCFG[pbs]           PBSALTIX=TRUE
RMCFG[pbs]           USEVNODES=TRUE
RMCFG[pbs]           IGNHNODES=TRUE

NODESETPOLICY        ONEOF 
NODESETATTRIBUTE     FEATURE 
NODESETISOPTIONAL    FALSE 
NODESETPRIORITYTYPE  MINLOSS
NODESETLIST          cnode1,cnode2,cnode3,cnode4
NODESETDELAY         1:00
NODESETMAXUSAGE      .5

NODEALLOCATIONPOLICY PRIORITY 
NODEACCESSPOLICY     SINGLEJOB

NODECFG[cnode1[1]]   FEATURES=cnode1
NODECFG[cnode1[2]]   FEATURES=cnode1
NODECFG[cnode1[3]]   FEATURES=cnode1
NODECFG[cnode1[4]]   FEATURES=cnode1

NODECFG[cnode2[1]]   FEATURES=cnode2
NODECFG[cnode2[2]]   FEATURES=cnode2
NODECFG[cnode2[3]]   FEATURES=cnode2
NODECFG[cnode2[4]]   FEATURES=cnode2

NODECFG[cnode2[1]]   FEATURES=cnode3
NODECFG[cnode2[2]]   FEATURES=cnode3
NODECFG[cnode2[3]]   FEATURES=cnode3
NODECFG[cnode2[4]]   FEATURES=cnode3

NODECFG[cnode2[1]]   FEATURES=cnode4
NODECFG[cnode2[2]]   FEATURES=cnode4
NODECFG[cnode2[3]]   FEATURES=cnode4
NODECFG[cnode2[4]]   FEATURES=cnode4

SRCFG[rsv1]          PERIOD=INFINITY
SRCFG[rsv1]          HOSTLIST=l:cnode1[1],cnode1[2],cnode1[3],cnode1[4]
SRCFG[rsv1]          CLASSLIST=admin
SRCFG[rsv1]          MAXJOB=1

An smp node is represented in Moab as a NODESET. Each node board in the node needs to be associated with a NODESET, which is done by setting the FEATURES on the node to the NODESET's name. NODESETPOLICY, NODESETATTRIBUTE, NODESETPRIORITYTYPE, and NODESETISOPTIONAL must all be set to the corresponding values indicated in the previous configuration for node board scheduling to work. NODESETDELAY is available for node board scheduling.

NODEALLOCATIONPOLICY must be set to PRIORITY.

Standing reservations can be used to group the virtual nodes of a node so that policies can set on the smp node. For example, MAXJOB=1 can be used to limit only one running job on the smp node at a time. When listing virtual nodes in the the standing reservation's HOSTLIST l: must be used.

Note: mdiag -n -v can be used to view the total statistics of each host node gathered by Moab.

compute node summary
Name                    State   Procs      Memory         Disk          Swap      Speed   Opsys   Arch Par   Load Rsv 

smp1[1]                  Idle    2:2     11264:11264       0:0           0:0       1.00       -      - loc   0.00   0 
smp1[2]                  Idle    2:2     11264:11264       0:0           0:0       1.00       -      - loc   0.00   0 
smp1[3]                  Idle    2:2     11264:11264       0:0           0:0       1.00       -      - loc   0.00   0 
smp2[1]               Running    1:2         0:4096        0:0           0:0       1.00       -      - loc   1.00   1 
smp2[2]                  Idle    2:2      4096:4096        0:0           0:0       1.00       -      - loc   0.00   0 
smp2[3]                  Idle    2:2      4096:4096        0:0           0:0       1.00       -      - loc   0.00   0 
smp3[1]               Running    1:4         0:8129        0:0           0:0       1.00       -      - loc   3.00   1 
smp3[2]                  Idle    4:4      8129:8129        0:0           0:0       1.00       -      - loc   0.00   0 
smp3[3]                  Idle    4:4      8129:8129        0:0           0:0       1.00       -      - loc   0.00   0 

smp summary
smp1                             6:6     33792:33792 
smp2                             4:6      8192:12288 
smp3                             8:12    16258:24387 

-----                     ---   20:24    58242:70467       0:0           0:0     

Total Nodes: 9  (Active: 2  Idle: 7  Down: 0)

Home Up Previous