Do you have CPU, network, and disk utilization data for 1, 2, 3, 4, and 5 tiles? It might be useful to help determine the bottleneck. I tend to use esxtop in batch mode.
Another way to look for bottlenecks is to compare your workload component scores with some of the ones published on the VMmark results page. A well-tuned tile should achieve an overall score 1.5-1.6 until CPU saturation. Most published results go a bit past saturation and the individual tiles scores are closer to 1.4. In your test, 4 tiles should achieve close to a score of 6 unless there is a bottleneck. That implies that at 4 tiles you already have a performance problem. I would suggest looking at each workload's normalized score and compare them with published numbers. That might tell you where to look. For instance, poor fileserver scores relative might point to storage.
If you post the tilescore.out, I can take a look.
Also, how is the storage configured - # of LUNs, RAID level, etc.?
