How Much Computation Power do you need for Near-Data Processing in Cloud ?

In modern cloud server systems, the big data application domain is one of the most important ones these days. However, as a large amount of data has to be transferred between storage unit and computing unit, the conventional server systems failed to fully utilize their resources. Near-data processing has been proposed to mitigate this inefficiency. Due to the reduced data transfer and its scalability, utilizing near-data processing can perform better than servers with high-end computing units. It can reduce data transfer by offloading computation to the embedded CPU in the storage device. In this work, based on the evaluation and analysis, we demonstrate advantages and disadvantages of existing server systems with near-data processing and show the potential of them with an alternative design. We also demonstrate that embedded CPUs are not enough to run applications such as big data applications. Instead, our evaluations show that computing units with better energy efficiency such as embedded GPU or ASIC can be a promising alternative that better exploits the advantage of near-data processing.