Multi-core benchmarks

Hi All,

I've been fixing and testing L3DT's mosaic system on multi-core processors, and I thought you might be interested to see some of the performance results.

The graph below shows the pixel throughput of the texture map calculation for single core, dual core, triple core and quad core systems. Basically, this graph shows how many pixels the texture generator can process in a given amount of time, so larger numbers mean faster calculations.

There are a few interesting things to note here:

1. When making a RAM-only texture (i.e. non-mosaic, shown as red bars), the throughput increases near-linearly with the number of cores. A quad-core machine works out as about 3.4x faster than a single core machine, which is a rather good result. This means that you can just keep throwing cores at L3DT Professional, and the performance will continue to improve. Based on these numbers, an 8-core machine should run at about 6-7x faster than a single core machine with the same clock-speed, at least for non-mosaic maps.
2. When making mosaic maps (shown as blue bars), the performance increases sub-linearly. This is due to all the cores sharing limited input/output bandwidth, namely the hard-disk. From the graph it is clear that the I/O bandwidth is starting to saturate at four cores, and adding more cores will not make a dramatic improvement to performance. However, adding faster mass storage (e.g. RAM drives) may make a substantially improvement to performance on mosaic map calculations with multi-core machines.

For completeness, the conditions for these tests were:

Also, for the mosaic calculations, all maps were mosaics (heightfield, water map, attributes map, normals map, light map, shadow map texture map). Similarly, for the RAM maps, all maps were RAM-only (i.e. non-mosaic).

That's all for today. If you'd like to make any comments or post questions, please do so in the users' forum.

Cheerio, Aaron.

PS: I you are using a multi-core processor with L3DT Professional, I would strongly recommend you upgrade to the latest developmental build (L3DT Pro v2.5d build 10, 12th of July '08), as it contains some substantial performance and stability improvements in the mosaic map system for multi-core builds.

Supplementary results

I re-ran the same benchmark tests for L3DT Professional in Ubuntu 8.04 with WINE 0.9.59. Plotted in the same fashion and on the same scale, these are the results for Linux+WINE:

Note that the same general trends apply, but that emulation via WINE decreases the speed by ~20% for mosaic maps and 20-40% for RAM-only maps when compared to Windows Vista.