Performance Evaluation of Container-based Virtualization for High Performance Computing Environments (1709.10140v1)

Abstract: Virtualization technologies have evolved along with the development of computational environments since virtualization offered needed features at that time such as isolation, accountability, resource allocation, resource fair sharing and so on. Novel processor technologies bring to commodity computers the possibility to emulate diverse environments where a wide range of computational scenarios can be run. Along with processors evolution, system developers have created different virtualization mechanisms where each new development enhanced the performance of previous virtualized environments. Recently, operating system-based virtualization technologies captured the attention of communities abroad (from industry to academy and research) because their important improvements on performance area. In this paper, the features of three container-based operating systems virtualization tools (LXC, Docker and Singularity) are presented. LXC, Docker, Singularity and bare metal are put under test through a customized single node HPL-Benchmark and a MPI-based application for the multi node testbed. Also the disk I/O performance, Memory (RAM) performance, Network bandwidth and GPU performance are tested for the COS technologies vs bare metal. Preliminary results and conclusions around them are presented and discussed.

• The paper demonstrates that container-based virtualization can achieve near-native HPC performance, with Docker incurring a 2.89% overhead and Singularity outperforming bare-metal by 5.42%.
• It employs rigorous benchmarks, including the HPL and GPU tests, to compare CPU, memory, disk I/O, and network metrics across LXC, Docker, and Singularity.
• The findings underscore the viability of containerization in HPC, with Singularity well-integrated and Docker showing potential for improvement in disk I/O efficiency.

Evaluation of Container-Based Virtualization in High-Performance Computing

The paper "Performance Evaluation of Container-based Virtualization for High Performance Computing Environments" by Carlos Arango, Rémy Dernat, and John Sanabria presents a detailed comparative analysis of three popular container-based OS virtualization technologies: Linux Containers (LXC), Docker, and Singularity. The primary focus of the research is the applicability of these technologies within High-Performance Computing (HPC) environments, examining their capacity to deliver isolated and resource-efficient computational contexts.

The research is structured around comprehensive benchmarking scenarios that assess the performance of LXC, Docker, and Singularity against traditional bare-metal environments across several key metrics: CPU speed, disk I/O performance, memory throughput, network bandwidth and latency, and GPU capabilities. A distinctive feature of this paper is its usage of industry standard HPC benchmarks such as the HPL-Benchmark and specialized GPU tests, which provide a robust basis for performance comparison.

Key Findings and Results

• CPU Performance: The paper reveals that Docker and Singularity are comparably close to bare-metal performance, with Docker showing a 2.89% overhead and Singularity outperforming bare-metal setup slightly by 5.42%. In contrast, LXC exhibited a 7.76% performance overhead.
• Disk I/O and Memory Performance: Docker displayed significant performance penalties in both sequential and random disk read/write operations, attributed to its AUFS layered filesystem. LXC and Singularity performed closer to native levels, with LXC having a slight edge in random I/O tasks. Memory bandwidth tests showed Docker to lag by approximately 36% in comparison to native performance.
• Network and GPU Performance: LXC achieved superior network bandwidth and latency results, while Docker's networking lagged due to complex packet processing layers. For GPU computations, Docker surprisingly yielded better performance than native setups—likely due to NVIDIA’s optimizations—while Singularity also performed well.

Implications for HPC and Future Directions

The findings underscore the viability of using containerization in HPC settings, particularly emphasizing Singularity due to its seamless integration with existing HPC infrastructures and minimal performance overhead. Enhancements in Docker's GPU handling suggest significant potential for machine learning and GPU-intensive tasks in containerized environments. However, Docker's I/O performance issues need addressing for its broader application in scientific computing.

This research contributes to our understanding of containerization within the HPC domain, suggesting that containerized environments can substantially meet the performance requirements traditionally reserved for bare-metal systems. The paper suggests that continued focus on optimizing these technologies—especially in areas of I/O and network performance—will be crucial.

In conclusion, the paper not only informs infrastructure architects of the efficacies and limitations associated with current container-based virtualization technologies for HPC but also encourages advancements tailored towards enhancing container performance in future research. This creates new possibilities for improved resource efficiency and flexibility in HPC environments, facilitating continued innovation in computational science.