The scheduler overhead and the cost of context-switches are vastly overstated compared to alternatives. The per thread stack space in effect has virtually no run-time cost, and starting off at a single 4k page for a stack, thousands still only waste a miniscule about of memory.