Efficient Compute at the Edge: Optimizing Energy Aware Data Structures for Emerging Edge Hardware

The advent of exascale computing, with the unparalleled rise in the scale of data in Internet of Things (IoT), high performance computing (HPC), and big data domains, both at the center and the edge of the system, requires optimal exploitation of energy-efficient computing hardware dedicated for edge processing. Emerging hardware for data processing at the edge must take advantage of advanced concurrent data locality-aware algorithms and data structures in order to provide better throughput and energy efficiency. Their design must be performance portable for their implementation to perform equally well on the edge hardware as well as other high performance computing, embedded and accelerator platforms. Concurrent search trees are one such widely used back-end for many important big data systems, databases, and file systems. We analyze DeltaTree, a concurrent energy-efficient and locality-aware data structure based on relaxed cache-oblivious model and van Emde Boas trees, on Intel's specialized computing platform Movidius Myriad 2, designed for machine vision and computing capabilities at the edge. We compare the throughput and energy efficiency of DeltaTree with B-link tree, a highly concurrent B+tree, on Movidius Myriad 2, along with a high performance computing platform (Intel Xeon), an ARM embedded platform, and an accelerator platform (Intel Xeon Phi). The results show that DeltaTree is performance portable, providing better energy-efficiency and throughput than B-link tree on these platforms for most workloads. For Movidius Myriad 2 in particular, DeltaTree performs really well with its throughput and efficiency up to 4× better than B-link tree.