This page contains discussion about advanced grid configurations. In other words how simulation nodes can be located in virtual space to produce optimal performance from both network and cpu consumption perspectives. Another important characteristics is the relative maximum size of objects compared to the simulation node sizes.

The following examples demonstrate how these characteristics vary depending on the chosen configuration. In practice these configurations may vary in different virtual locations of the grid depending on the requirements and importance of the volume in question.

If you know other good configurations or existing articles please refer them here or extend this article. Critique is very much welcome as well.

We will use the following values in the performance requirement calculations:

• Observation Frame Size: 64 bytes
• Observation Frequency: 1/s
• Bandwidth Available for Each Simulation Node: 100 Mbits/s
• Bandwidth Available for Each Simulation Client: 250 kbits/s
• Bandwidth Reserved for Node to Node Communication: 25 Mbits/s
• Node Observation Distance to Outside Node Volume: Node Width / 2
• Maximum Object Size: Node Observation Distance / 2 = Node Width / 4

• Maximum Clients per Node: 300 (See: Performance)
• Maximum Observer Objects per Client: 400 (See: Performance)

Simple Cubic Lattice

Simplest grid configuration is simple cubic lattice as illustrated in the following picture in 2D:

Simple cubic configuration does not provide redundancy or local load balancing. Local load balancing means that the same space is simulated by several simulations which exchange state information. One object is simulated by one simulation and one client is connected to one simulation at a time. All the servers need to simulate each object with simple 3d physics model to predict movement but the internal state is simulated only by one simulation. Different simulations may be run by different service providers and they may have different implementations. Local load balancing offers networking benefits as clients can be directed to different simulations. For example client could choose to join the simulation with lowest ping or one with well compliant implementation.

The same configuration illustrated in 3D:

Bandwidth Considerations for Simple Cubic Lattice

• Number of Total Observing Nodes: 26
• Number of Effective Observing Nodes (per object): 7
• Effective Node to Node Observation Bandwidth per Object: 448 bytes / s
• Maximum Number of Objects per Node: 25 Mbits/s / 448 bytes / s = 3550
• Maximum Number of Objects per Cube Volume: 5580
• Maximum Number of Clients per Cube Volume: 300

Dual Cubic Lattice

Dual cubic lattice consists of two intersecting simple cubic lattices as illustrated in the following 2D picture:

Dual cubic configuration offers spatial load balancing, local load balancing and redundancy. In redundant grid one simulation node may shutdown and the grid remains accessible in that location. Whether the objects in the simulation which was shutdown still exist after the event depends on succesful migration of the objects to the other simulation nodes governing that volume. Dual cubic lattice also minimizes the amount of offline volume down to 1/8th of node volume when two overlapping nodes shutdown simultaneously.

Bandwidth Considerations for Dual Cubic Lattice

• Number of Total Observing Nodes: 26 + 8 = 35
• Number of Effective Observing Nodes (per object): 11
• Effective Node to Node Observation Bandwidth per Object: 704 bytes / s
• Maximum Number of Objects per Node: 3550
• Maximum Number of Objects per Cube Volume: 2*3550 = 7100
• Maximum Number of Clients per Cube Volume: 2*300 = 600