Upper Bounds on Number of Steals in Rooted Trees

Abstract Inspired by applications in parallel computing, we analyze the setting of work stealing in multithreaded computations. We obtain tight upper bounds on the number of steals when the computation can be modeled by rooted trees. In particular, we show that if the computation with n processors starts with one processor having a complete k-ary tree of height h (and the remaining n−1 processors having nothing), the maximum possible number of steals is ∑ni=1(k−1)i(hi)