Current scheduler walks the entire TestConfig list on every scheduling epoch. This was observed to be a bottleneck on large machines: the scheduler is unable to feed the entire machine with jobs. We can do much better by trying each scheduling class exactly once.