Oguzhan Burak Tekin

I fully redesigned and reinstalled an HPC in operation, adding future-proof functionality with almost no downtime impacting ongoing user projects during expansion. Changes were addressed to provide all management functions in a highly available setup and increase users' HPC experience. While transitioning the HPC to a new OS release to support brand new versions of HPC applications, older applications that are still required but not supported on the OS release are also kept accessible with the use of purposefully prepared containers.

An end-to-end HPC system that I designed and implemented for CFD Analysis with Ansys of critical components of turbine engines. I used Slurm as a schedular, implemented BeeGFS parallel files system running on SSD and NVMe technology for user space and scratch, respectively, and integrated additional SAN storage space for automatic D2D backups.

Two different petabyte scale Lustre parallel file system-based storage, a total of nine PB, were implemented for the HPC system running mainly Schlumberger Seismic Research Software and Omega used by the national petroleum and natural gas mining research company. An out-of-the-box, bulletproof design that was beyond the requirements was made, and extensive tests were done for data integrity in the harshest conditions. Up to 60 GB/s throughput was reached in data operations on the Lustre system.

An HPC cluster that I built for interdisciplinary academic research at Sabanci University, Istanbul. I led the project to redesign and implement the HPC system based on the OpenHPC schemas.

To use the existing hardware effectively, I created a primarily virtual environment on Proxmox VE and selected an operating system based on Rocky Linux 8.6. I used technologies like Open vSwitch, SR-IOV, and VF with Infiniband HCAs, bonding, GPU computing, Ceph, BeeGFS, and BeeOND. I also implemented NVMe caching on storage tiers for improved IO performance and a metadata service with DRBD HA and pacemaker, going beyond regular networking and hardware boundaries. For the deployment of compute nodes and other functions, I used xCAT. Finally, I equipped the system with the latest versions of GNU, Intel, PGI Compilers, Python, R, OpenMPI stack, Slurm, Environment Modules, and Singularity. The system is monitored with dashboards in Grafana and XDMoD.

Researchers have access to the system over a security gateway with SSH or RDP. Now, they can use the latest applications compiled from open-source code, containers, and commercial apps and visualize data on the HPC system without moving data back to their local environments.

HPC infrastructure is redesigned comprehensively and reinstalled to create a unified cluster of Intel- and AMD-based hardware to serve researchers in the METU Physics Department in field of condensed matter, astrophysics, and other computational physics areas.

The client requested an analysis of their HPC system to identify root causes that extended the time required to run HPC jobs. To understand HPC system performance, I implemented a dashboard to analyze throughput, IOPS, and metadata performance, ran thorough tests, and worked on necessary steps toward improvements.

The client was a prominent children's cancer research hospital running their research applications on Azure. They were using a 24-node BeeGFS parallel file system which was underperforming on Azure.

I reviewed configurations of the BeeGFS cluster and implemented necessary dashboards to monitor the storage system and identify the root cause of underperforming. After running the required performance tests, the root cause was identified. The server resources for the metadata system were not enough and caused saturation. The necessary advisory was provided to improve the performance long term.

As a solution consultant, I established the video surveillance of the International Port of Turkmenbashi. I designed the infrastructure of 12 data centers for video surveilling of an area of 1,200,000 square meters with more than 4,000 units of 8K cameras in QCIF format. When the project was signed, I organized the shipment of hardware, which consisted of more than 170 servers, 40 conventional storage systems, and 10PB parallel distributed storage to a fictive warehouse in Istanbul with a Lodz-Tilburg-Istanbul route. All of them were installed there to the last screw and then transferred from Istanbul to the final destination 3,500 km away in just four weeks to keep up with the project's deadline for a great opening, which leaders of multiple countries will attend.

A fully redesigned HPC deployment for Koc University, one of the most prominent private universities in Türkiye, to fulfill the emerging needs of researchers. HPC management functions are fully designed with high availability in mind, whereas very high file system performance is obtained through a multimodal BeeGFS parallel file system using top-notch NVIDIA InfiniBand switches. The system is being monitored for multiple aspects with Grafana, Zabbix, and XDMod.