|
U sers are encouraged to submit
jobs requiring 8–16 CPUs in the underutilized queues of com_nhec,
csl_nhec, com_nh and csl_nh in the blackforest complex.
Jobs submitted in these queues will be scheduled on one of our two16-CPU
Nighthawk nodes.
Due to larger communication bandwidth and lower latencies over shared
memory 8–16-task, we expect these 8–16 CPU jobs jobs to run
faster on Nighthawk nodes than on the equivalent 2–4 Winterhawk
nodes. However, users must change their LoadLeveler scripts to specify:
#@ network.MPI = css,shared,ip
#@ environment = MP_SHARED_MEMORY=true
For details, please see:
http://www.scd.ucar.edu/docs/ibm/ref/comm.html
|
