GPFS: A Shared-Disk File System for Large Computing Clusters
FAST '02 Proceedings of the Conference on File and Storage Technologies
Fault tolerant high performance computing by a coding approach
Proceedings of the tenth ACM SIGPLAN symposium on Principles and practice of parallel programming
Scalable diskless checkpointing for large parallel systems
Scalable diskless checkpointing for large parallel systems
High Performance Linux Clusters: With OSCAR, Rocks, openMosix, and MPI (Nutshell Handbooks)
High Performance Linux Clusters: With OSCAR, Rocks, openMosix, and MPI (Nutshell Handbooks)
Improving the write performance of an NFS server
WTEC'94 Proceedings of the USENIX Winter 1994 Technical Conference on USENIX Winter 1994 Technical Conference
A graphical interface for private cloud and cluster management
Proceedings of the 51st ACM Southeast Conference
| Hi-index | 0.00 |
In recent years, significant research has been conducted to boost the performance and increase the reliability of high performance computing (HPC) clusters. As the number of compute nodes in modern HPC clusters continues to grow, it is critical to design clusters with low power consumption and low failure rate. In particular, it is widely known that the internal disk drives of compute nodes (in the case of diskfull clusters) are a major source of failures. In addition, these diskfull HPC clusters tend to require more power and cooling requirements compared to diskless clusters. In this paper, we propose and implement a large-scale Infiniband-based diskless HPC cluster. The paper presents the cluster configuration and evaluates its performance using various High Performance LINPACK (HPL) benchmarks. The performance is measured in terms of the overall efficiency, speed in Giga-Floating Point Operations per Second (GFLOPS), and HPL execution time. We also measure temperature and power consumption. We compare the performance measurements of our diskless cluster to its diskfull counterpart. For our measurement and comparison, we consider three cluster sizes of 32, 64, and 126 compute nodes.