Balanced Allocations推荐.pdfVIP

CHAPTER FOURTEEN * Balanced Allocations In thi s chapter we examine a s mple yet powerful var ant of the classic balls-and-bins paradigm , with applications to hashing and dynamic resource allocation . 1 4 . 1 . The Power of Two Choices Suppose that we sequential ly place n balls into n bins by putting each ball into a bin chosen independently and uniformly at random. We studied this classic balls-and-bins problem in Chapter 5. There we showed that , at the end of the process, the most balls in any bin - the maximum load is 8 ( In n I In In n ) with high probability. In a variant of the process, each ball comes with d possible destination b ns, each chosen independently and uniformly at random , and is placed in the least full bin among the d possible locations at the time of the placement. The orig nal balls-and­ bins process corresponds to the case where d 1 . Surpr si ngly, even when d = 2 , the behavior is completely different: when the process terminates, the maximum load is In In n /In 2 + 0( I ) with high probability. Thus. an apparently m nor change in the ran­ dom allocation process results in an exponential decrease in the maximum load. We may then ask what happens if each ball has three choices; perhaps the resulting load is then 0 ( In In In n ) . We shalJ consider the general case of d choices per balJ and show that , when d :=_ 2 . with high probability the maximum load is ln ln n / ln d + E-) ( 1 ) . Although having more than two cho ces does reduce the maximum load, the reduction changes it by only a constant factor. so it remains () ( ln ln n ) . 14.1.1. The Upper Bound Theorem 14. 1 : Suppose that n balls are sequentially placed into n bins in the follow­ ing manner. For each ball, d :=_ 2 bins are chosen independently and uniformly at random (