Energy and performance impact of aggressive volunteer computing with multi-core computers

The rapid advances in multi-core architecture and the predicted emergence of 100-core personal computers bring new appeal to volunteer computing. The availability of massive compute power under-utilized by personal computing tasks is a blessing to volunteer computing customers. Meanwhile the reduced performance impact of running a foreign workload, thanks to the increased hardware parallelism, makes volunteering resources more acceptable to PC owners. In addition, we suspect that with aggressive volunteer computing, which assigns foreign tasks to active computers (as opposed to idle ones in the common practice), we can obtain significant energy savings. In this paper, we assess the efficacy of such aggressive volunteer computing model by evaluating the energy saving and performance impact of co-executing resource-intensive foreign workloads with native personal computing tasks. Our results from executing 30 native-foreign workload combinations suggest that aggressive volunteer computing can achieve an average energy saving of around 52% compared to running the foreign workloads on high-end cluster nodes, and around 33% compared to using the traditional, more conservative volunteer computing model. We have also observed highly varied performance interference behavior between the workloads, and evaluated the effectiveness of foreign workload intensity throttling.