I saw that you mentioned running Hadoop/Spark on your cluster? Do you have any comments on that? I just made a 3 node RPi3 cluster and I have to admit, it's really slow. Like getting the spark Scala interpreter running takes about a minute before anything I type registers in the console. (NOte: this is my own cluster not a PicoCluster TM)

When all 4 cores are under load, the BCM2837 on the Raspberry Pi 3 will rapidly reduce the clock rate after about 60 seconds (from 1.2 GHz to 600 MHz, or even 300 MHz if the core temperature exceeds 85 degrees Celsius). Cooling is required if you want decent performance for prolonged periods of time.

The RevPi Core 3 is equipped with a high performance cooler which is able to sustain 1.2 GHz at full load for about 20 minutes (at room temperature). After that the clock rate will be moderately reduced to 1.15 GHz.

A 4% heat throttle seems an unusually subtle adjustment. If testing has found it to be effective that is a good finesse.