Real-Time Analytics as the Killer Application for Processing-In-Memory

Abstract — While Processing-In-Memory (PIM) has been widely researched for the last two decades, it was never truly adopted by the industry and remains mostly within the academic research realm. This is mainly because (1) in-memory compute engines were too slow, and (2) a real-world application that could really benefit from PIM was never identified. In recent years, the first argument became untenable, but the community still struggles to identify a poster-child application that refutes the second argument. In this position paper, we argue that real-time analytics is the killer application the community has been searching for. We show that several inherent characteristics of real-time analytics fit perfectly within the PIM paradigm, and identify Redis – a popular in-memory key-value framework – as a convenient platform to showcase PIM in this domain. Finally, we call for a targeted hardware-software co-design effort towards a PIM acceleration architecture for real-time analytics. I