Estimating reliability of workers for cooperative distributed computing

Internet supercomputing is an approach to solving partitionable, computation-intensive problems by harnessing the power of a vast number of interconnected computers. Forthe problem of using network supercomputing to perform a large collection of independent tasks, prior work introduced a decentralized approach and provided randomized synchronousalgorithms that perform all tasks correctly with high probability, while dealing with misbehaving or crash-prone processors. The main weaknesses of existing algorithms is that they assume either that the average probability of a non-crashed processor returningincorrect results is inferior to 1/2, or that the probability of returning incorrect results is known to each processor. Here we present a randomized synchronous distributed algorithm that tightly estimates the probability of each processor returning correct results. Starting with the set P of n processors, let F be the set of processors that crash. Our algorithm estimates the probability p-i of returning a correct result for each processor i \in P - F, making the estimates available to all these processors. The estimation is based on the (e, d)-approximation, where each estimated probability pi of p-i obeys the bound Pr[pi(1 - e) = pi=pi(1+e)]>1 - d, for any constants d>0 and e>0 chosen by the user. An important aspect of this algorithm is that each processor terminates without global coordination. We assess the efficiency of the algorithm in three adversarial models as follows. For the model where the number of non-crashed processors |P - F | is linearly bounded the time complexity T(n) of the algorithm is O(log n), work complexity W(n) is O(n log n), and message complexity M(n) is O(n log2 n). For the model where |P - F | is bounded by a fractional polynomial we have T(n) = O(n1 - a logn loglogn), W(n) = O(nlogn loglogn), and M(n) = O(nlog2n loglogn). For the model where |P - F| is bounded by a poly-logarithm we have T(n) = O(n), W(n) = O(n polylog n), and M(n) = O(n log2 n polylog n). All bounds are shown to hold with high probability.