1) Test 1 : In this test, the Kixum Benchmark scene is rendered by a small renderfarm, using the standard 5-nodes configuration included with Carrara 5 Pro 5.03. Table 1 on the right shows the speed gain as each new PC is added. In this test all the PCs are relatively close as far as computing power is concerned.
In the screenshot, we see that the 'N' tiles do not represent 'Nodes' (where 1 node = 1 PC) but 'Computing Units' (CU). A Pentium 4 with HT (HyperThreading) is seen as 2 CUs or 2 gray 'N's. The same can be said of a single CPU dual core PC (A64 X2), or a dual CPU single core computer. A dual Xeon with HT or a dual Opteron dual core is 4 CU = 4 N.
9 N = 3 x P4 with HT (6 N)
+ 1 dual CPU Athlon MP (2 N)
+ 1 single core Athlon 64 (1 N)
2) Test 2 : In Test 2, nodes 4 & 5 are faster than the others. We can see that even though a Dual Xeon 2.8 is 43 % faster than a P4-3.3, the render time with 4 nodes is about the same as in Test 1. Carrara's default tile size (128) is not optimal when PCs of different speeds are used simultaneously.
13 N = 1 P4 (2N) + 1 single core Athlon 64 (1N) + 1 dual CPU Athlon MP (2N) + 1 Dual Xeon (4N) + 1 Dual Opteron 275 (4N)
3) Test 3 : Here, the server (an overclocked Dual Opteron 275) is 3.67 times faster than the only node (an Athlon XP 2.2 GHz). The total render time is longer than with the Dual Opteron alone ! It is therefore not recommended to use very fast and very slow PCs together in the same renderfarm.
4) Suggestion to improve C5 Pro tile rendering algorithm
The image on the right is from Test 3. We can see that the N node is stuck on a difficult tile, while the server has already finished its work and is idle. To improve the efficiency, the last tile of a render should be assigned to _all_ available computers (server and nodes) in the network. That way, the entire renderfarm would not have to wait for its slowest member to finish.