Democratization of HPC cloud services with automated parallel solvers and application containers

摘要

In this paper, we investigate several design choices for HPC services at different layers ofthe cloud computing architecture to simplify and broaden its use cases. We start with theplatform-as-a-service (PaaS) layer and compare direct and iterative parallel linear equationsolvers. We observe that several matrix properties that can be identified before startinglong-running solvers can help HPC services automatically select the amount of computingresources per job, such that the job latency is minimized and the overall job throughput is maximized.As a proof of concept, we use classical problems in structural mechanics and mesh theseproblems with increasing granularities leading to various matrix sizes, ie, largest having 1 billionnon-zero elements. In addition tomatrix size,wetake into account matrix condition numbers, preconditioningeffects, and solver types and execute these finite element analysis (FEA) over an IBMHPCcluster.Next, we focus on the infrastructure-as-a-service (IaaS) layer and exploreHPC applicationperformance, load isolation, and deployment issues using application containers (Docker)while also comparing them to physical and virtualmachines (VM) over a public cloud.