Teodor Nicolau

I worked on many projects that involved managing complete deployments of HPC clusters ranging from 64 to 4,000 nodes, with or without GPUs (Nvidia, AMD), with only a list of media access control addresses and info from the customer about the operating system and what IP ranges to use for Ethernet, InfiniBand, and Baseboard Management Controller. I handled complete health checks and configuration of all cluster components. This also involved consistency checks of all firmware components, upgrades, and BIOS configuration and tuning. The success of these deployments was measured through benchmarks, where the results were within the agreed standard deviation.

I was involved in many projects requiring different software installations and technologies, including Slurm Workload Manager, GPUDirect RDMA and GPUDirect Storage (GDS), ROCm, Cray programming environment, custom Lmod modules, Spack, Open MPI, Intel MPI, and OpenSM.

The project involved managing the installation of a Lustre file system on Hewlett Packard Enterprise (HPE) hardware on standard Red Hat High Availability (HA). The solution leveraged ZFS with Grafana and Prometheus for monitoring. The success of the project depended on seamlessly integrating the Lustre file system with the customer's existing environment, with IOR as a benchmark used to evaluate the performance of the Lustre file system installation.