Cryptonight is one of the possible base algorithms for cryptocurrencies. It belongs to the group of memory-bound algorithms, designed to prevent mining on specialized processors and ASICs by using 2MB of memory for each hash. Thus, it is not easy to adapt for parallel computing. The aim of this work is to prove theoretically and experimentally that this algorithm can still be optimized for a specialized multicore processor to make mining more energetically efficient than on CPU. This article describes the process of optimization, which was conducted using the following methods: data clustering, storage of repeatedly used data in local memory, usage of SIMD for parallel computing, data prefetch. Those methods are first explained, their supposed effectiveness analyzed, and then implemented. As a result, two schemes of algorithm optimization were created: first one is based on the usage of MALTs slave cores, which compute hashes independently. Although memory-boundness creates multiple problems, we were able to increase the efficiency by clustering data. The second scheme is more complicated, it suggests using SIMD processors for most cryptographic computations and also involves data prefetch, which becomes possible if more than one hash is calculated on one core at the same time. All the results are demonstrated in the paper and they indicate that it is indeed possible to optimize Cryptonight for a specialized multicore processor MALT. The practical results show that energy efficiency has increased 5 times in comparison with CPU.
Energy-efficient data prefetch buffering for low-end embedded processors
