Speaker
Description
COKA (Computing On Kepler Architecture) is a small HPC cluster, designed for codes prototyping and benchmarking, installed in 2015 at INFN and University of Ferrara, with a DP peak performance of about 100 TFLOP/s. The cluster is made up on GPU-dense nodes, hosting each: 8 x NVIDIA K80, which are Dual-GPUs boards based on the Kepler architecture; 2 x Haswell CPUs; and 2 x 56Gb/s Infiniband cards. Like the Voyager, it outlived its original mission and it has been continuously operational since almost ten years. It is still used nowadays to prepare, test and setup HPC applications, AI pipelines, as well as for educational purposes. In this talk, after a brief introduction on the cluster infrastructure, we focus on lessons learnt during these years, along with the plans for upcoming hardware and software upgrades.
