Visualization System VIS5D, Cluster Control

from a linear to a logarithmic scale, change a colour palette, or something like that; simple changes like that are easily made directly to the generated file. EMPOST’s calling syntax is

empost orders_file results_file

The second argument is the name of a file that EMPOST produces, containing numerical results of the named orders, e.g. integrals, fluxes, mode matching coefficients, and so on. These are in the form of assignment statements, and are parsed by POEMS when postprocessing is completed. POEMS then uses these to compute the value of the penalty function for the current iteration.

2.5. The Visualization System: VIS5D

VIS5D is an advanced visualization program originally written for meteorological data. It runs under the X Window System, which is native to Linux and other Unix derivatives, but which has to be added to Windows and OS/2. Windows users can install Hummingbird Exceed, which works well with VIS5D once all the arcane X parameters are set up. See Appendix B for a working sample X configuration.

The Vis5D conversion code in empost is based on a stand-alone program by Theodore G. van Kessel. In this release, it is fully integrated into EMPOST. Both animated and static Vis5D files are generated using the MOVIE3D statement.

2.6. Cluster Control

There are lots of ways to structure a cluster, lots of communications styles (such as the Message Passing Interface (MPI), and lots of cluster management systems such as the Sun Grid Engine (SGE). POEMS is not tied to any of these, but is easily adapted to them. The main script runs on a frontend machine. Inter-host communication requires no specific support other than a high capacity, low latency TCP/IP network. FIDO uses TCP/IP socket communication between fido subdomains running on different hosts, and local communication between subdomains running on the same host. To allow the user control, cluster control is not hardcoded into POEMS, but relies on an external script. The supplied script is fidossh, which uses a shared file system (e.g. NFS, XFS, or PVFS2) for communication and ssh for cluster node control, This design is suitable for clusters of up to perhaps 20 nodes, depending on filesystem performance. The high bandwidth host-to- host communication is organized in a distributed fashion between cluster hosts, so the frontend machine does not become a bottleneck in small and medium sized clusters. For larger clusters, FIDO can use a hierarchical supervision scheme, where a single frontend node is not forced to supervise hundreds or thousands of hosts, but the cluster script would have to be tailored for the application.

The downside of this flexibility is that the user has to apportion the work manually. Future versions of POEMS will help automate this, based on the CPU speed, number of cores, and amount of memory possessed by each host. Probably it will remain semiautomatic unless the subdomains can be made very small.