Multi-resource Fair Allocation with Bounded Number of Tasks in Cloud Computing Systems (1410.1255v3)

Abstract: Dominant resource fairness (DRF) is a popular mechanism for multi-resource allocation in cloud computing systems. In this paper, we consider a problem of multi-resource fair allocation with bounded number of tasks. Firstly, we propose the lexicographically max-min normalized share (LMMNS) fair allocation mechanism, which is a natural generalization of DRF, and design a non-trivial optimal algorithm to find a LMMNS fair allocation, whose running time is linear in the number of users. Secondly, we prove that LMMNS satisfies envy-freeness (EF) and group strategy-proofness (GSP), and analysis the approximation ratios of LMMNS, by exploiting the properties of the optimal solution. Thirdly, we propose a modified version of LMMNS, which is the second mechanism satisfying sharing incentive, EF, and GSP. Finally, we have implemented LMMNS, and show that it has a good average-case performance, especially when the number of resources is 2.