cHPCe: AN EFFICIENT RESOURCE PROVISIONING ON CONTAINERIZED HPC ENVIRONMENT

Abstract

High Performance Computing (HPC) systems are changing the way computation is performed and reproduced without sacrificing raw performance using Container technologies compared to hypervisor-assisted virtualization technologies. It primarily supports continuously evolving data-intensive applications such as computational fluid dynamics, seismic tomography, molecular biology, and Proteomics. Despite having several existing Container solutions for HPC, users still need to be convinced about the suitability of Container orchestration solutions for their extreme-scale applications due to the lack of thorough performance assessment on recent advances. HPC hardware architectures and programming models continuously evolve, but the platform models suffer from the HPC community’s awareness. Hardware and application-aware co-design with power-awareness have recently attracted the most to empower the platform models. In this thesis, to address these problems, we design and develop a unique containerized HPC environment (cHPCe) from scratch using Linux namespaces. First, we provide an in-depth performance analysis of the cHPCe using x86 and OpenPOWER systems on subcomponents such as processor, memory, interconnect, and IO. The performance of the cHPCe is compared with BareMetals and VMs using the benchmarks HPC Challenge (HPCC) and IOZone. Our experimental results achieve 0.13% less compute performance penalty at its peak performance on cHPCe compared to the BareMetal-based solution for x86 systems. In contrast, a VMbased solution introduces an overhead of 20%. Moreover, we observe inconsistent behavior for memory performance with a worst-case penalty of 9.68% compared to achieved peak performance. However, similar behavior is reported for cHPCe with an overhead of less than 3% and 2% in the worst case for the latency and bandwidth, respectively, compared to the BareMetal for network and disk performance. Further, this work proposes an analytical model for data locality and memory bandwidth contention-aware Container placement strategy for our developed cHPCe to mitigate resource ix interference by co-hosted applications. Performance is evaluated and compared against LXD, Docker Swarm, Kubernetes, and Singularity using HPCC and NAS parallel (NPB) benchmarks. The experimental results show that data locality and memory bandwidth contention-awareness reduce the overall execution time of the benchmark in cHPCe by 51.41% in the best case compared to the Docker Swarm. Most Dynamic Power Management (DPM) approaches proposed for the HPC environment are based on profile-guided power-performance prediction techniques. However, the complexity of DPM approaches in a multi-tenant containerized HPC environment (cHPCe) increases significantly due to the varying demands of users and the contention of shared resources. Moreover, there is limited research into software-level monitoring of power consumption. The proposed study in this thesis aims to present a real-time hybrid power-performance prediction approach using a Long Short-Term Memory (LSTM) machine learning model with the rolling update mechanism. It also proposes a power-cap determination framework with resource contention awareness to fine-tune real-time power consumption at the thread level. The proposed power-aware cHPCe (pcHPCe) is evaluated and compared with native BareMetal execution using the NPB and HPCC applications. Our experimental results show that the power-performance prediction model achieves an accuracy of 91.39% on average in real-time with an overhead of 1.6% of the total computing power per node. Our resource contention-aware power-cap selection framework attains significant power savings up to 13.1%.