When HPC meets Big Data in the Cloud

Abstract

High Performance Computing (HPC) has traditionally been seen as a specialist area where most of its applications are scientific applications that require large CPU capabilities (e.g., at petaFLOPS scale). With the advent of cloud computing and the new paradigm of “everything-imaginable-as-a-service”, scientists and researchers are able to deploy their HPC applications in the cloud. Entering the era of “Big Data’, HPC has rapidly evolved in ways that are substantially different from its past. Essentially, the GHz war is over. Processor architectures go for coprocessors (or accelerators) as exemplified by NVIDIA’s GPGPUs and Intel’s Xeon Phi; or continue with the self-hosted CPUs by increasing the number of cores instead of clock rate (e.g., Tilera’s 100-core processor, Intel’s 48-core SCC and 72-core “Knights Landing”). Several implications are with these new design requirements and architectural changes, including the importance of exploiting the data locality, the new parallel programming models to deliver superior performance with ease, and how to harness the new hardware’s massive parallelism. In this talk, I will introduce HKU’s eXCloud project that implemented a handful of migration techniques working at diverse granularities to achieve better data locality for high-performance cloud computing. I will also report preliminary results of our Crocodiles project which realizes the concept of Cloud-on-Chip (CoC) for supporting power-efficient Big Data computing on a single many-core chip.