The CMX kernel that Erik recommended has the best performance (based on my research several years ago) when preemptive multitasking is required. It had the lowest interrupt latency. However, the documentation was not so good, but maybe its better nowadays.

If your requirements are for a small footprint and you can get by with cooperative multitasking, Salvo is quite good (see http://www.pumpkininc.com). Its configurability is outstanding, allowing you to enable only the features you need. Salvo's documentation is superb and has a good tutorial section. A freeware version is available if you just want to experiment. The freeware version is provided in library form without source code for obvious reasons.