Clustered FDTD

Brief Description

The Cluster Versions of CrystalWave and OmniSim provide scalable compute power for the FDTD engine of these products. This allows you to solve problems that are too big to run on a single PC and also to solve problems more quickly.

Features

• Familiar Windows based Graphical User Interface. All control is via a familiar CrystalWave or OmniSim user interface – once set up all cluster control occurs transparently to the user.
• Linux/x86 and Windows 2000/XP cluster nodes.
• 32 bit Linux and Windows nodes - 3GB per cluster node (Windows XP†), 2GB per node (Linux-32), 3.5GB per nodes (Linux64 or Win64 OS).
• 64 bit Linux nodes - virtually unlimited memory.
• Both 1D and 2D domain tiling (division of 3D problem into smaller elements)
• Cluster Management Interface (CMI) – all cluster management occurs via the CMI, including installation of nodes, diagnostics, resource limiting.
• Automatic load balancing – giving slower nodes less work to do.
• Support for SMP – each node can have multiple CPUs and/or multi-core CPUs.
• Network: 100MB/s or Gigabit Ethernet.
• Up to 60 compute nodes (approximate) supported. (Future versions will support more).
• Supports Intel Pentium 4, Pentium4-EE, Pentium-D, Pentium-M, Intel Core family, Xeon, AMD Athlon64 and AMD Opteron.
• Almost all features of the single-CPU versions of CrystalWave and OmniSim FDTD engines are available also on cluster version.

Performance

Photon Design’s 3D-FDTD engine has constantly outperformed competing products, usually by a significant margin. The exact performance obtained under cluster operation is dependent on several things:

• network bandwidth
• domain shape
• domain size – larger domains will run closer to the ideal speed.

As a guideline you can typically expect a problem using at least 1GB per node to run better than 80% of the ideal speed – i.e. the speed that would be obtained if all inter-node communication was infinitely fast. This assumes that your network is quiet except for clustering traffic – additional network traffic will slow down the cluster.

Windows Cluster

The Windows cluster is designed as a convenient ad-hoc clustering setup where unused PCs on your network can be conveniently added to a cluster calculation with little added work. The setup is shown below.

Linux Cluster

The Linux version of the cluster FDTD is designed for more dedicated high-performance cluster use. In this scenario typically the compute nodes are on a private network, inaccessible from designers’ desktop PCs and accessible only through a “head node”. This configuration is shown below.

Hardware Recommendations

If you do not yet have a cluster and would like some advice on what to buy to obtain the maximum performance with our CrystalWave and OmniSim Cluster Editions, see our document fdtd_cluster_hardware_recommendations.pdf for recommendations on what to look for in a computer cluster.