Some of these algorithms (e.g., for processor scheduling) have per-decision costs comparable to those of non-market mechanisms in current use; others have costs that are much greater. In general, these costs will be acceptable for objects of sufficient size and processes of sufficient duration. The question of the appropriate scale at which to apply market mechanisms can be addressed by additional study but will best be addressed by experience in actual computational markets. The proposals made here can doubtless be improved upon; they are merely intended to illustrate some of the issues involved in incentive engineering for computational markets, and to provide a starting point for discussion and design. Any advances toward lower costs, greater effectiveness, and better incentive structures will shift tradeoff points in favor of finer-grained application of market mechanisms.
Even heavy overhead costs would leave intact a solid case for market mechanisms in computation. This case rests on the value of doing the right thing (or something like it) with some overhead costs, rather than doing something blatantly wrong with polished efficiency. And when finding the right thing to do requires cooperation, competition, and freewheeling experimentation, the value of decentralized systems with market accountability becomes very great indeed.