Exhaustive key search on clusters of GPUs

Exhaustive search is generally a last resort for solving a problem: each possible state of a system is generated and evaluated against a condition to find if the problem solution is attained. In some cases, for example in the reversal of cryptographic hash functions that make use of the salting technique, there are very few valid alternatives. However, the set of candidate solutions can be extremely large and therefore very substantial computing resources are needed to walk through the search space in a reasonable time. On the other hand, exhaustive search is very often embarrassingly parallel and so the task can be easily accelerated by distributing the work on a multitude of devices. In this paper we propose a pattern to parallelize general exhaustive searches on a heterogeneous and hierarchical network of computing nodes. We validate this pattern by applying it to the reversal of MD5 and SHA1 hash functions, both at coarse grain (work dispatching among nodes) and at fine grain (work made by each thread), reaching linear scalability with increasing computing power of the participating nodes. In particular, we show how to implement and optimize the hash key search on a GPGPU, achieving near-maximal throughput on various models of NVIDIA devices programmed with CUDA