Well, no, we don't need to accept that cost, it's just a constraint for a concurrency model to choose, i.e. maximizing locality. Apart from that another big one for massively multicore systems is minimizing cross-core synchronization. You can meet both, just not with threads as your programming model.