I made a suggestion that exec time local balancing should be

performing a kind of round robin algorithm. The idea is that

looking at the current loads of all nodes the kernel can calculate a

percentage of execs to put into each node. It would make sense

that the least noded nodes get the bigger percentage and the most

loaded get the least percentage. There may be thresholds which

exclude very loaded nodes. At this point I haven't come up with

the exact math put here are some examples:

Unloaded cluster - round-robin to keep load balanced

Node Load percentage

1 200 25%

2 200 25%

3 200 25%

4 200 25%

One overloaded node (skipped) two other nodes that are the same

- try to bring others evenly up

Node Load percentage

1 2000 0%

2 200 50%

3 200 50%

One overloaded node (skipped) one remaining node double the

others load - weight more heavily to least loaded node

Node Load percentage

1 2000 0%

2 400 33%

3 200 66%