NCEP CCS Conversion Guide |
Snow/Frost System Overview |
Back to Main Page |
Snow/Frost Configuration
The CCS computer for this phase of the acquisition consists of two seperate POWER4 Clusters (SPs) called Snow and Frost. Each cluster contains 22 Regatta Servers. The servers are tied together with two colony switch planes. The first two Regatta servers on each system are reserved for interactive use, GPFS Servers and LoadLeveler Class 1 batch processing. The remaining servers are compute servers. Each Regatta server is divided into 8 logical partitions (LPARs). NFS is served up by two HACMP servers outside of each cluster.
Regatta Server Characteristics
Improving Performance
#@network.MPI = csss,shared,us
The "csss" specification will stripe messages across both switch "planes".
Improving Turnaround
#@ blocking = unlimited
This will give you faster turnaround by using any free initiators.
#@ wall_clock_limit = hh:mm:ss
This will specify the smallest wall time needed to complete the job.
#@ network.MPI = csss,shared,us
The "shared" specification will allow to share a node with another
job giving your job a higher probability of initiating sooner.
Dedicated Access to a Node
#@ node_usage = not_shared
If you have less than 4 MPI tasks on a node this will prevent LoadLeveler
from assigning other work, either serial jobs or other MPI tasks, to the
nodes you are using. For everyday use of the "dev" queue, it is not
recommended you specify this in your command file. If you are doing
timing tests and desire the least amount of variability, then you should
specify this in your command file.
#@ network.MPI = csss,not_shared,us
This approach will not prevent serial jobs from initiating on the node
your job is running.
Back to Main Page |